Systems and methods for selecting a macroblock mode in a video encoder

[0001] This application claims the benefit under 35 U.S.C. §119(e) of U.S. Provisional Application Ser. No. 60/403,851, filed Aug. 14, 2002, the entirety of which is hereby incorporated by reference. This application is also related to U.S. patent application Ser. No. 10/642,107, entitled SYSTEMS AND METHODS FOR ADAPTIVELY FILTERING DISCRETE COSINE TRANSFORM (DCT) COEFFICIENTS IN A VIDEO ENCODER, filed on the same date and commonly owned with the present application, the entirety of which is hereby incorporated by reference herein.

[0002] 1. Field of the Invention

[0003] The invention generally relates to video encoding techniques. In particular, the invention relates to mode selection for macroblocks.

[0004] 2. Description of the Related Art

[0005] A variety of digital video compression techniques have arisen to transmit or to store a video signal with a lower data rate or with less storage space. Such video compression techniques include international standards, such as H.261, H.263, H.263+, H.263++, H.264, MPEG-1, MPEG-2, MPEG-4, and MPEG-7. These compression techniques achieve relatively high compression ratios by discrete cosine transform (DCT) techniques and motion compensation (MC) techniques, among others. Such video compression techniques permit video data streams to be efficiently carried across a variety of digital networks, such as wireless cellular telephony networks, computer networks, cable networks, via satellite, and the like, and to be efficiently stored on storage mediums such as hard disks, optical disks, Video Compact Discs (VCDs), digital video discs (DVDs), and the like. The encoded data streams are decoded by a video decoder that is compatible with the syntax of the encoded data stream.

[0006] For relatively high image quality, video encoding can consume a relatively large amount of data. However, the communication networks that carry the video data can limit the data rate that is available for encoding. For example, a data channel in a direct broadcast satellite (DBS) system or a data channel in a digital cable television network typically carries data at a relatively constant bit rate (CBR) for a programming channel. In addition, a storage medium, such as the storage capacity of a disk, can also place a constraint on the number of bits available to encode images.

[0007] As a result, a video encoding process often trades off image quality against the number of bits used to compress the images. Moreover, video encoding can be relatively complex. For example, where implemented in software, the video encoding process can consume relatively many CPU cycles. Further, the time constraints applied to an encoding process when video is encoded in real time can limit the complexity with which encoding is performed, thereby limiting the picture quality that can be attained.

[0008] One conventional method for rate control and quantization control for an encoding process is described in Chapter 10 of Test Model 5 (TM5) from the MPEG Software Simulation Group (MSSG). TM5 suffers from a number of shortcomings. An example of such a shortcoming is that TM5 does not guarantee compliance with the Video Buffer Verifier (VBV) requirement. As a result, overrunning and underrunning of a decoder buffer can occur, which undesirably results in the freezing of a sequence of pictures and the loss of data.

[0009] The invention is related to methods and apparatus that provide improved video encoding. Conventional video encoding techniques inadequately determine the encoding mode for a macroblock. One embodiment according to the invention advantageously includes the coding cost for motion vectors into the mode decision, thereby enhancing the quality of the selection of the mode. One embodiment further uses an L1 norm in the decision, thereby advantageously reducing the computational complexity of taking the coding cost for motion vectors into the mode decision.

[0010] One embodiment of the invention is a method of selecting a mode for the encoding of a macroblock in a video encoder, the method including: performing a motion search to select a motion vector; determining a residual error for the motion vector with a plurality of macroblock modes; estimating a coding cost from motion vectors for at least some of the plurality of macroblock modes; and selecting the mode for the encoding of the macroblock based on both the residual error and the coding cost associated with the mode.

[0011] One embodiment of the invention is a method of selecting a mode for the encoding of a macroblock (MB) in a video encoder, the method including using both an indication for residual error and a coding cost for a motion vector for a mode to determine which mode is selected for the macroblock.

[0012] One embodiment of the invention is a video encoder that is configured to select a mode for the encoding of a macroblock (MB), the video encoder including means for using both an indication for residual error and a coding cost for a motion vector for a mode to determine which mode is selected for the macroblock.

[0013] One embodiment of the invention is a computer program embodied in a tangible medium including a module with instructions for selecting a mode for the encoding of a macroblock (MB), the computer program including instructions for using both an indication for residual error and a coding cost for a motion vector for a mode to determine which mode is selected for the macroblock.

[0014] One embodiment of the invention is a circuit configured to select a mode for the encoding of a macroblock in a rate and quality efficient manner, the circuit including: a circuit configured to perform a motion search to select a motion vector; a circuit configured to determine a residual error for the motion vector with a plurality of macroblock modes; a circuit configured to estimate a coding cost associated with the motion vectors for at least some of the plurality of macroblock modes; and a circuit configured to select the mode for the encoding of the macroblock based on both the residual error and the coding cost associated with the mode. In one embodiment, the circuit is embodied in a video encoder.

[0037] Although this invention will be described in terms of certain preferred embodiments, other embodiments that are apparent to those of ordinary skill in the art, including embodiments that do not provide all of the benefits and features set forth herein, are also within the scope of this invention. Accordingly, the scope of the invention is defined only by reference to the appended claims.

[0038] FIG. 1 illustrates a sequence of pictures 102. While embodiments of the invention are described in the context of MPEG-2 and pictures, the principles and advantages described herein are also applicable to other video standards including H.261, H.263, H.264, MPEG-1, and MPEG-4, as well as video standards yet to be developed. The term “picture” will be used herein and encompasses pictures, images, frames, visual object planes (VOPs), and the like. A video sequence includes multiple video images usually taken at periodic intervals. The rate at which the pictures of frames are displayed is referred to as the picture rate or frame rate. The pictures in a sequence of pictures can correspond to either interlaced images or to non-interlaced images, i.e., progressive images. In an interlaced image, each image is made of two separate fields, which are interlaced together to create the image. No such interlacing is performed in a non-interlaced or progressive image.

[0039] The sequence of pictures 102 can correspond to a movie or other presentation. It will be understood that the sequence of pictures 102 can be of finite duration, such as with a movie, or can be of unbound duration, such as for a media channel in a direct broadcast satellite (DBS) system. An example of a direct broadcast satellite (DBS) system is known as DIRECTV®. As shown in FIG. 1, the pictures in the sequence of pictures 102 are grouped into units known as groups of pictures such as the illustrated first group of pictures 104. A first picture 106 of the first group of pictures 104 corresponds to an I-picture. The other pictures in the group of pictures can correspond to P-pictures or to B-pictures.

[0040] In MPEG-2, a picture is further divided into smaller units known as macroblocks. It will be understood that in other video standards, such as MPEG-4, a picture can be further divided into other units, such as visual object planes (VOPs). Returning now to MPEG-2, an I-picture is a picture in which all macroblocks are intra coded, such that an image can be constructed without data from another picture. A P-picture is a picture in which all the macroblocks are either intra coded or forward predictively coded. The macroblocks for a P-picture can be encoded or decoded based on data for the picture itself, i.e., intra coded, or based on data from a picture that is earlier in the sequence of pictures, i.e., forward predictively coded. A B-picture is a picture in which the macroblocks can be intra coded, forward predictively coded, backward predictively coded, or a combination of forward and backward predictively coded, i.e., interpolated. During an encoding and/or a decoding process for a sequence of pictures, the B-pictures will typically be encoded and/or decoded after surrounding I-pictures and/or P-pictures are encoded and/or decoded. An advantage of using predictively-coded macroblocks over intra-coded macroblocks is that the number of bits used to encode predictively-coded macroblocks can be dramatically less than the number of bits used to encode intra-coded macroblocks.

[0041] The macroblocks include sections for storing luminance (brightness) components and sections for storing chrominance (color) components. It will be understood by one of ordinary skill in the art that the video data stream can also include corresponding audio information, which is also encoded and decoded.

[0042] FIG. 2 illustrates an example of an encoding environment in which an embodiment of the invention can be used. A source for unencoded video 202 provides the unencoded video as an input to an encoder 204. The source for unencoded video 202 can be embodied by a vast range of devices, such as, but not limited to, video cameras, sampled video tape, sampled films, computer-generated sources, and the like. The source for unencoded video 202 can even include a decoder that decodes encoded video data. The source for unencoded video 202 can be external to the encoder 204 or can be incorporated in the same hardware as the encoder 204. In another example, the source for unencoded video 202 is a receiver for analog broadcast TV signals that samples the analog images for storage in a digital video recorder, such as a set-top box known as TiVo®.

[0043] The encoder 204 can also be embodied in a variety of forms. For example, the encoder 204 can be embodied by dedicated hardware, such as in an application specific integrated circuit (ASIC), by software executing in dedicated hardware, or by software executing in a general-purpose computer. The software can include instructions that are embodied in a tangible medium, such as a hard disk or optical disk. In addition, the encoder 204 can be used with other encoders to provide multiple encoded channels for use in direct broadcast satellite (DBS) systems, digital cable networks, and the like. For example, the encoded output of the encoder 204 is provided as an input to a server 206 together with the encoded outputs of other encoders as illustrated in FIG. 2. The server 206 can be used to store the encoded sequence in mass storage 208, in optical disks such as a DVD 210 for DVD authoring applications, Video CD (VCD), and the like. The server 206 can also provide the data from the encoded sequence to a decoder via an uplink 212 to a satellite 214 for a direct broadcast satellite (DBS) system, to the Internet 216 for streaming of the encoded sequence to remote users, and the like. It will be understood that an encoded sequence can be distributed in a variety of other mediums including local area networks (LANs), other types of wide area networks (WANs), wireless networks, terrestrial digital broadcasts of television signals, cellular telephone networks, dial-up networks, peer-to-peer networks, and the like. In one embodiment, the encoder 204 encodes the sequence of pictures in real time. In another embodiment, the encoder 204 encodes the sequence of pictures asynchronously. Other environments in which the encoder 204 can be incorporated include digital video recorders, digital video cameras, dedicated hardware video encoders and the like.

[0044] FIG. 3 illustrates an example of decoding environments, which include decoder buffers that are modeled during the encoding process by a Video Buffer Verifier (VBV) buffer. An encoded sequence of pictures can be decoded and viewed in a wide variety of environments. Such environments include reception of direct broadcast satellite (DBS) signals via satellite dishes 302 and set top boxes, playback by digital video recorders, playback through a DVD player 304, reception of terrestrial digital broadcasts, and the like. For example, a television set 306 can be used to view the images, but it will be understood that a variety of display devices can be used.

[0045] For example, a personal computer 308, a laptop computer 310, a cell phone 312, and the like can also be used to view the encoded images. In one embodiment, these devices are configured to receive the video images via the Internet 216. The Internet 216 can be accessed via a variety of networks, such as wired networks and wireless networks.

[0046] FIG. 4 is a block diagram that generally illustrates the relationship between an encoder 402, an encoder buffer 404, a decoder 406, a decoder buffer 408, and a constant-bit-rate data channel 410. In another embodiment, the bit rate of the constant-bit-rate data channel can vary slightly from channel-to-channel depending on a dynamic allocation of data rates among multiplexed data channels. For the purposes of this application, this nearly constant bit rate with a slight variation in data rate that can occur as a result of a dynamic allocation of data rate among multiplexed data channels will be considered as a constant bit rate. For example, the encoder 402 can correspond to an encoder for a programming channel in a direct broadcast satellite (DBS) system, and the decoder 406 can correspond to a decoder in a set-top box that receives direct broadcast satellite (DBS) signals. The skilled practitioner will appreciate that the data rate of the constant-bit-rate data channel 410 for actual video data may be less than the data rate of the constant-bit-rate data channel 410 itself because some of the actual transmission data may be occupied for overhead purposes, such as for error correction and for packaging of data. The skilled practitioner will appreciate that the methods described herein are directly applicable to constant-bit-rate encoding, as described in the MPEG standard document, but also to variable-bit-rate encoding. For the case of variable bit-rate, the transmission bit rate can be described in terms of a long-term average over a time period that can be a few seconds, a few minutes, a few hours, or any other suitable time-interval, together with a maximal bit rate that can be used to provide data to a decoder buffer. Data can be provided from the channel to the decoder buffer at the maximal bit rate until the decoder buffer is full; at that point, the data channel waits for decoding of the next picture, which will remove some data from the decoder buffer, and then transfer of data from the channel to the decoder buffer resumes. The term “bit rate” used hereafter can be either some constant bit rate or a long-term average of variable bit rate encoding. In one embodiment of a constant bit rate encoder, the encoder produces a data stream with a relatively constant bit rate over a group of pictures.

[0047] For streaming applications such as a direct broadcast satellite (DBS) system or for recording of live broadcasts such as in a home digital video recorder, the encoder 402 receives and encodes the video images in real time. The output of the encoder 402 can correspond to a variable bit rate (VBR) output 412. The variable bit rate (VBR) output 412 of the encoder 402 is temporarily stored in the encoder buffer 404. A function of the encoder buffer 404 and the decoder buffer 408 is to hold data temporarily such that data can be stored and retrieved at different data rates. It should be noted that the encoder buffer 404 and the decoder buffer 408 do not need to be matched, and that the encoder buffer 404 is a different buffer than a video buffer verifier (VBV) buffer, which is used by the encoder 402 to model the occupancy of the decoder buffer 408 during the encoding process.

[0048] The encoder buffer 404 can be implemented in dedicated memory or can be efficiently implemented by sharing system memory, such as the existing system memory of a personal computer. Where the memory used for the encoder buffer 404 is shared, the encoder buffer 404 can be termed a “virtual buffer.” It will be understood that larger memories, such as mass storage, can also be used to store video data streams and portions thereof.

[0049] The encoder buffer 404 buffers the relatively short-term fluctuations of the variable bit rate (VBR) output 412 of the encoder 402 such that the encoded data can be provided to the decoder 406 via the constant-bit-rate data channel 410. Similarly, the decoder buffer 408 can be used to receive the encoded data at the relatively constant bit rate of the constant-bit-rate data channel 410 and provide the encoded data to the decoder 406 as needed, which can be at a variable bit rate. The decoder buffer 408 can also be implemented in dedicated memory or in a shared memory, such as the system memory of a personal computer. Where implemented in a shared memory, the decoder buffer 408 can also correspond to a virtual buffer.

[0050] The MPEG standards specify a size for the decoder buffer 408. The size of the decoder buffer 408 is specified such that an MPEG-compliant data stream can be reliably decoded by a standard decoder. In the MPEG-2 standard, which for example is used in the encoding of a DVD, the buffer size specified is about 224 kB. In the MPEG-1 standard, which for example is used in the encoding of a video compact disc (VCD), the buffer size is specified to be about 40 kB. It will be understood by one of ordinary skill in the art that the actual size of the encoder buffer 404 and/or the decoder buffer 408 can be determined by a hardware designer or by a software developer by varying from the standard.

[0051] Although it will be understood that the actual size of the decoder buffer 408 can vary from standard, there exist practical limitations that affect the size and occupancy of the decoder buffer 408. When the size of the decoder buffer 408 is increased, this can correspondingly increase the delay encountered when a sequence is selected and playback is initiated. For example, when a user changes the channel of a direct broadcast satellite (DBS) set-top box or skips forwards or backwards while viewing a DVD, the retrieved data is stored in the decoder buffer 408 before it is retrieved by the decoder 406 for playback. When the decoder buffer 408 is of a relatively large size, this can result in an infuriatingly long delay between selection of a sequence and playback of the sequence. Moreover, as will be described later in connection with FIG. 5, the encoded data can specify when playback is to commence, such that playback can begin before the decoder buffer 408 is completely full of data.

[0052] In one embodiment, playback of a sequence begins upon the earlier of two conditions. A first condition is a time specified by the MPEG data stream. A parameter that is carried in the MPEG data stream known as vbv-delay provides an indication of the length of time that data for a sequence should be buffered in the decoder buffer 408 before the initiation of playback by the decoder 406. The vbv-delay parameter corresponds to a 16-bit number that ranges from 0 to 65,535. The value for the vbv-delay parameter is counted down by the decoder 406 by a 90 kHz clock signal such that the amount of time delay specified by the vbv-delay parameter corresponds to the value divided by 90,000. For example, the maximum value for the vbv-delay of 65,535 thereby corresponds to a time delay of about 728 milliseconds (mS). It will be understood that the vbv-delay can initiate playback of the sequence at a time other than when the decoder buffer 408 is full so that even if the decoder buffer 408 is relatively large, the occupancy of the decoder buffer 408 can be relatively low.

[0053] A second condition corresponds to the filling of the decoder buffer 408. It will be understood that if data continues to be provided to the decoder buffer 408 after the decoder buffer 408 has filled and has not been emptied, that some of the data stored in the decoder buffer 408 will typically be lost. To prevent the loss of data, the decoder 406 can initiate playback at a time earlier than the time specified by the vbv-delay parameter. For example, when the size of the decoder buffer 408 corresponds to the specified 224 kB buffer size, bit-rates that exceed 2.52 Mega bits per second (Mbps) can fill the decoder buffer 408 in less time than the maximum time delay specified by the vbv-delay parameter.

[0054] The concept of the VBV buffer in the MPEG specification is intended to constrain the MPEG data stream such that decoding of the data stream does not result in an underrun or an overrun of the decoder buffer 408. It will be understood that the VBV buffer model does not have to be an actual buffer and does not actually have to store data. However, despite the existence of the VBV buffer concept, the video encoding techniques taught in MPEG's Test Model 5 (TM5) do not guarantee VBV compliance, and buffer underrun and overrun can occur.

[0055] Buffer underrun of the decoder buffer 408 occurs when the decoder buffer 408 runs out of data. This can occur when the bit rate of the constant-bit-rate data channel 410 is less than the bit rate at which data is consumed by the decoder 406 for a relatively long period of time. This occurs when the encoder 402 has used too many bits to encode the sequence relative to a specified bit rate. A visible artifact of buffer underrunning in the decoder buffer 408 is a temporary freeze in the sequence of pictures.

[0056] Buffer overrun of the decoder buffer 408 occurs when the decoder buffer 408 receives more data than it can store. This can occur when the bit rate of the constant-bit-rate data channel 410 exceeds the bit rate consumed by the decoder 406 for a relatively long period of time. This occurs when the encoder 402 has used too few bits to encode the sequence relative to the specified bit rate. As a result, the decoder buffer 408 is unable to store all of the data that is provided from the constant-bit-rate data channel 410, which can result in a loss of data. This type of buffer overrun can be prevented by “bit stuffing,” which is the sending of data that is not used by the decoder 406 so that the number of bits used by the decoder 406 matches with the number of bits sent by the constant-bit-rate data channel 410 over a relatively long period of time. However, bit stuffing can introduce other problems as described in greater detail later in connection with FIGS. 9A and 9B.

[0057] The VBV buffer model concept is used by the encoder 402 in an attempt to produce a video data stream that will preferably not result in buffer underrun or overrun in the decoder buffer 408. In one embodiment, the occupancy levels of the VBV buffer model are monitored to produce a video data stream that does not result in buffer underrun or overrun in the decoder buffer 408. It should be noted that overrun and underrun in the encoder buffer 404 and in the decoder buffer 408 are not the same. For example, the conditions that result in a buffer underrun in the decoder buffer 408, i.e., an encoded bit rate that exceeds the bit rate of the constant-bit-rate data channel 410 for a sustained period of time, can also result in buffer overrun in the encoder buffer 404. Further, the conditions that result in a buffer overrun in the decoder buffer 408, i.e., an encoded bit rate that is surpassed by the bit rate of the constant-bit-rate data channel 410 for a sustained period of time, can also result in a buffer underrun in the encoder buffer 404.

[0058] FIG. 5 is a chart that generally illustrates decoder buffer occupancy as data is provided to a decoder buffer at a constant bit rate while data is consumed by a decoder at a variable bit rate. In a conventional system based on MPEG TM5, the data stream provided to the decoder disadvantageously does not guarantee that the decoder buffer is prevented from buffer underrun or overrun conditions. In the illustrated example, the data is provided to the decoder buffer at a constant bit rate and the decoder uses the data to display the video in real time.

[0059] Time (t) 502 is indicated along a horizontal axis. Increasing time is indicated towards the right. Decoder buffer occupancy 504 is indicated along a vertical axis. In the beginning, the decoder buffer is empty. A maximum level for the buffer is represented by a B_MAX 528 level. An encoder desirably produces a data stream that maintains the data in the buffer below the B_MAX 528 level and above an empty level. For example, the decoder buffer can be flushed in response to a skip within a program, in response to changing the selected channel in a direct broadcast satellite (DBS) system or in a digital cable television network, and the like. The decoder monitors the received data for a system clock reference (SCR), as indicated by SCR(0) 506. The system clock reference (SCR) is a time stamp for a reference clock that is embedded into the bit stream by the encoder and is used by the decoder to synchronize time with the time stamps for video information that are also embedded in the bit stream. The time stamps indicate when video information should be decoded, indicate when the video should be displayed, and also permit the synchronization of visual and audio samples.

[0060] An example of a picture type pattern that is commonly used in real-time video encoding is a presentation order with a repeating pattern of IBBPBBPBBPBBPBB. Despite the fact that I-pictures consume relatively large amounts of data, the periodic use of I-pictures is helpful for example, to permit a picture to be displayed in a relatively short period of time after a channel change in a DBS system.

[0061] The picture presentation or display order can vary from the picture encoding and decoding order. B-pictures depend on surrounding I- or P-pictures and not from other B-pictures, so that I- or P-pictures occurring after a B-picture in a presentation order will often be encoded, transmitted, and decoded prior to the encoding, transmitting, and decoding of the B-picture. For example, the relatively small portion of the sequence illustrated in FIG. 5 includes data for pictures in the order of IPBBP, as a P-picture from which the B-pictures depend is typically encoded and decoded prior to the encoding and decoding of the B-pictures, even though the pictures may be displayed in an order of IBBPBBPBBPBBPBB. It will be understood that audio data in the video presentation will typically not be ordered out of sequence. Table I summarizes the activity of the decoder with respect to time. For clarity, the illustrated GOP will be described as having only the IPBBP pictures and it will be understood that GOPs will typically include more than the five pictures described in connection with FIG. 5.

[0062]

	TABLE I
	time	activity
	<T0	data accumulates in the buffer
	T0	I-picture is decoded
	T1	I-picture is presented, first P-picture is decoded
	T2	first B-picture is decoded and presented
	T3	second B-picture is decoded and presented
	T4	first P-picture is presented, second P-picture is
		decoded

[0063] In one embodiment, the decoder buffer ignores data until a picture header with a presentation time stamp (PTS) for an I-frame is detected. This time is indicated by a time TTS₀(0) 508 in FIG. 5. This bypassing of data prevents the buffering of data for part of a picture or frame or the buffering of data that cannot be decoded by itself. After the time TTS₀(0) 508, the decoder buffer begins to accumulate data as indicated by the ramp R₀ 510.

[0064] For a time period τ₀(0) 512, the decoder buffer accumulates the data before the decoder begins using the data. This time period τ₀(0) 512 is also known as a pre-loading delay. Along the top of FIG. 5 are references for time that are spaced approximately evenly apart with a picture period equal to the inverse of the frame rate or inverse of the picture rate (1/R_f) 514. As will be described later, the location in time for the pictures can be indicated by time stamps for the corresponding pictures. At a time T₀ 516, the decoder retrieves an amount of data corresponding to the first picture of a group of pictures (GOP), which is an I-picture. The data stream specifies the time to decode the I-picture in a decoding time stamp (DTS), which is shown as a time stamp DTS₀(0) 518 and specifies the time T₀ 516.

[0065] The retrieval of data corresponding to the I-picture is indicated by the relatively sharp decrease 520 in decoder buffer occupancy. For clarity, the extraction of data from the decoder buffer is drawn in FIG. 5 as occurring instantaneously, but it will be understood by one of ordinary skill in the art that a relatively small amount of time can be used to retrieve the data. Typically, I-pictures will consume a relatively large amount of data, P-pictures will consume a relatively smaller amount of data, and B-pictures will consume a relatively small amount of data. However, the skilled practitioner will appreciate that intra macroblocks, which consume a relatively large amount of data, can be present in P-pictures and in B-pictures, as well as in I-pictures, such that P-pictures and B-pictures can also consume relatively large amounts of data. The I-picture that is decoded at the time T₀ 516 is not yet displayed at the time T₀ 516, as a presentation time stamp PTS₀(1) 522 specifies presentation at a time T₁ 524.

[0066] At the time T₁ 524, the decoder displays the picture corresponding to the I-picture that was decoded at the time T₀ 516. The time period PTS_OFFSET 526 illustrates the delay from the start of accumulating data in the decoder buffer for the selected sequence to the presentation of the first picture. A decoding time stamp DTS₀(1) 530 instructs the decoder to decode the first P-picture in the sequence at the time T₁ 524. The extraction of data from the decoder buffer is illustrated by a decrease 532 in buffer occupancy. In between the time T₀ 516 to the time T₁ 524, the decoder buffer accumulates additional data as shown by a ramp 534. A presentation time stamp PTS₀(4) 536 instructs the decoder to display the first P-picture at a time T₄ 538. In this example, the first P-picture is decoded earlier than it is presented such that the B-pictures, which can include backward predictively, forward predictively, or even bi-directionally predictively coded macroblocks, can be decoded.

[0067] At a time T₂ 540, the decoder decodes and displays the first B-picture as specified by a presentation time stamp PTS₀(2) 542. No decoding time stamp (DTS) is present because both the decoding and presenting occur at the same time period. It will be understood that in actual decoders, there can be a relatively small delay between the decoding and the displaying to account for computation time and other latencies. The amount of data that is typically used by a B-picture is relatively small as illustrated by a relatively small decrease 550 in decoder buffer occupancy for the first B-picture. It will be understood, however, that B-pictures can also include intra macroblocks that can consume a relatively large amount of data.

[0068] At a time T₃ 546, the decoder decodes and displays the second B-picture as specified by a presentation time stamp PTS₀(3) 548.

[0069] At the time T₄ 538, the decoder displays the first P-picture that was originally decoded at the time T₁ 524. At the time T₄ 538, the decoder also decodes a second P-picture as specified by the second P-picture's decoding time stamp DTS₀(4) 554. The second P-picture will be presented at a later time, as specified by a presentation time stamp (not shown). The decoder continues to decode and to present other pictures. For example, at a time T₅ 544, the decoder may decode and present a B-frame, depending on what is specified by the data stream.

Rate Control and Quantization Control Process

[0070] FIG. 6 is a flowchart that generally illustrates a rate control and quantization control process in a video encoder. It will be appreciated by the skilled practitioner that the illustrated process can be modified in a variety of ways without departing from the spirit and scope of the invention. For example, in another embodiment, various portions of the illustrated process can be combined, can be rearranged in an alternate sequence, can be removed, and the like. In another embodiment, selected portions of the illustrated process are replaced with processes from a rate control and quantization control process as disclosed in Chapter 10 of Test Model 5. The rate at which bits are consumed to encode pictures affects the occupancy of the decoder buffer during encoding. As illustrated by brackets in FIG. 6, portions of the process are related to bit allocation, to rate control, and to adaptive quantization. Bit allocation relates to estimating the number of bits that should be used to encode the picture to be encoded. Rate control relates to determining the reference quantization parameter Q_j that should be used to encode a macroblock. Adaptive quantization relates to analyzing the spatial activity in the macroblocks in order to modify the reference quantization parameter Q_j and calculate the value of the quantization parameter mquant_j that is used to quantize a macroblock.

[0071] The process begins at a state 602, where the process receives its first group of pictures. It will be understood that in one embodiment, the process may retrieve only a portion of the first group of pictures in the state 602 and retrieve remaining portions of the first group of pictures later. In the illustrated process, the pictures are grouped into groups of pictures before the pictures are processed by the rate control and quantization control process. A group of pictures starts with an I-picture and can include other pictures. Typically, but not necessarily, the other pictures in the group of pictures are related to the I-picture. The process advances from the state 602 to a state 604.

[0072] In the state 604, the process receives the mode or type of encoding that is to be applied to the pictures in the group of pictures. In the illustrated rate control and quantization control process, the decision as to which mode or type of encoding is to be used for each picture in the group of pictures is made before the pictures are processed by the rate control and quantization control process. For example, the group of pictures described earlier in connection with FIG. 5 have types IPBBP. The process advances from the state 604 to a state 606.

[0073] In the state 606, the process determines the number of P-pictures N_p and the number of B-pictures N_b in the group of pictures to be encoded. For example, in the group of pictures with types IPBBP, there are two P-pictures and there are two B-pictures to be encoded such that a value for N_p is 2 and a value for N_b is also 2. There is no need to track the number of I-pictures remaining, as the only I-picture in a group of pictures is the first picture. The process advances from the state 606 to a state 608.

[0074] In the state 608, the process initializes values for complexity estimators X_i, X_p, and X_b and for the remaining number of bits R allocated to the group of pictures that is to be encoded. In one embodiment, the process initializes the values for the complexity estimators X_i, X_p, and X_b according to Equations 1-3. $\begin{array}{cc}{X}_i=\frac{160·\mathrm{bit\_rate}}{115}& \left(\mathrm{Eq}.1\right)\\ X_p=\frac{160·\mathrm{bit\_rate}}{115}& \left(\mathrm{Eq}.2\right)\\ X_b=\frac{42·\mathrm{bit\_rate}}{115}& \left(\mathrm{Eq}.3\right)\end{array}$

[0075] In Equations 1-3, the variable bit_rate corresponds to the relatively constant bit rate (in bits per second) of the data channel, such as the constant-bit-rate data channel 410 described earlier in connection with FIG. 4. In another embodiment, bit_rate corresponds to the average or desired average bit rate of a variable bit rate channel. In yet another embodiment, bit_rate corresponds to a piece-wise constant bit rate value of a variable bit rate channel.

[0076] In one embodiment, the initial value R₀ for the remaining number of bits R at the start of the sequence, i.e., the initial value of R before encoding of the first group of pictures, is expressed in Equation 4 as R₀. At the start of the sequence, there is no previous group of pictures and as a result, there is no carryover in the remaining number of bits from a previous group of pictures. Further updates to the value for the remaining number of bits R will be described later in connection with Equations 27 and 28. $\begin{array}{cc}{R}_0=G& \left(\mathrm{Eq}.4\right)\\ G=\frac{\mathrm{bit\_rate}·N}{\mathrm{picture\_rate}}& \left(\mathrm{Eq}.5\right)\end{array}$

[0077] The variable G represents the number of bits that can be transferred by the data channel in an amount of time corresponding to the length of the presentation time for the group of pictures. This amount of time varies with the number of pictures in the group of pictures. In Equation 5, the variable bit_rate is in bits per second, the value of N corresponds to the number of pictures in the group of pictures (of all types), and the variable picture_rate is in pictures or frames per second. The process then advances from the state 608 to a state 610.

[0078] In the state 610, the process calculates an initial target number of bits T_i, T_p, or T_b, i.e., an initial target bit allocation, for the picture that is to be encoded. It should be noted that the pictures in a group of pictures will typically be encoded out of sequence when B-pictures are encoded. In one embodiment, the rate control and quantization control process calculates the initial target bit allocation for the picture according to the equation from Equations 6-8 for the corresponding picture type that is to be encoded. $\begin{array}{cc}{T}_i=\max \left\{\left(\frac{R}{(1+\frac{N_p{X}_p}{X_i{K}_p}+\frac{N_b{X}_b}{X_i{K}_b}}\right),\left(\frac{\mathrm{bit\_rate}}{8·\mathrm{picture\_rate}}\right)\right\}& \left(\mathrm{Eq}.6\right)\\ T_p=\max \left\{\left(\frac{R}{\left(N_p+\frac{N_b{K}_p{X}_b}{K_b{X}_p}\right)}\right),\left(\frac{\mathrm{bit\_rate}}{8·\mathrm{picture\_rate}}\right)\right\}& \left(\mathrm{Eq}.7\right)\\ T_b=\max \left\{\frac{R}{\left(N_b+\frac{N_p{K}_b{X}_p}{K_p{X}_b}\right)},\frac{\mathrm{bit\_rate}}{8·\mathrm{picture\_rate}}\right\}& \left(\mathrm{Eq}.8\right)\end{array}$

[0079] In Equation 6, T_i corresponds to the target bit allocation for the next picture to be encoded when the picture is the I-picture that starts a group of pictures, and T_i is determined by the higher of the two expressions in the brackets. In Equation 7, T_p corresponds to the target bit allocation for the next picture to be encoded when the next picture is a P-picture. In Equation 8, T_b corresponds to the target bit allocation for the picture when the picture is a B-picture. The values of the “universal constants” K_p and K_b depend on the quantization matrices that are used to encode the pictures. It will be understood that the values for K_p and K_b can vary. In one embodiment, the values for K_p and K_b are 1.0 and 1.4, respectively. In another embodiment, the value of these constants can be changed according to the characteristics of the encoded pictures, such as amount and type of motion, texture, color and image detail.

[0080] In one embodiment of the rate control and quantization control process, the process further adjusts the target bit allocation T_(i,p,b) from the initial target bit allocation depending on the projected buffer occupancy of the decoder buffer as will be described in greater detail later in connection with FIG. 7.

[0081] When the process has determined the target bit allocation for the next picture to be encoded, the process advances from the state 610 to a state 612. Also, the bits allocated to a picture are further allocated among the macroblocks of the picture. This macroblock bit allocation can be calculated by conventional techniques, such as techniques described in TM5, or by the techniques described herein in greater detail later in connection with a state 614. In addition, various orders or sequences in which a picture can advantageously be processed when encoded into macroblocks will be described in greater detail later in connection with FIGS. 8A and 8B.

[0082] In the state 612, the process sets initial values for virtual buffer fullness. In one embodiment, there is a virtual buffer for each picture type. The variables d_jⁱ, d_j^p, and d_j^b represent the virtual buffer fullness for I-pictures, for P-pictures, and for B-pictures, respectively. The variable j represents the number of the macroblock that is being encoded and starts at a value of 1. A value of 0 for j represents the initial condition. The virtual buffer fullness, i.e., the values of d_jⁱ, d_j^p, and d_j^b, correspond to the virtual buffer fullness prior to encoding the j-th macroblock such that the virtual buffer fullness corresponds to the fullness at macroblock (j−1). $\begin{array}{cc}{d}_0^i=10·\frac{r}{31}& \left(\mathrm{Eq}.9\right)\\ d_0^p=K_p·d_0^i& \left(\mathrm{Eq}.10\right)\\ d_0^b=K_b·d_0^i& \left(\mathrm{Eq}.11\right)\end{array}$

[0083] One example of a computation for the value of the reaction parameter r that appears in Equation 9 is expressed by Equation 12. It will be understood by one of ordinary skill in the art that other formulas for the calculation of the reaction parameter r can also be used. $\begin{array}{cc}r=2·\frac{\mathrm{bit\_rate}}{\mathrm{picture\_rate}}& \left(\mathrm{Eq}.12\right)\end{array}$

[0084] With respect to Equations 10 and 11, K_p and K_b correspond to the “universal constants” described earlier in connection with Equations 6-8. The process can advance from the state 612 to the state 614 or can skip to a state 616 as will be described in connection with the state 614.

[0085] The process can also include a macroblock mode decision 622 to determine whether a macroblock should be encoded as Intra, Inter_No_MC, and Inter_MC, and the like. It will be understood that the types modes that are available can depend on a variety of factors, such as the type of picture, whether the picture is interlaced or progressive. The macroblock mode decision can be executed in parallel or in series with the states 616, 619, 621. One embodiment of the macroblock mode decision that advantageously uses a coding cost for motion vectors in the mode decision will be described in greater detail later in connection with FIG. 17. The macroblock mode decision is provided as an input to the state 623.

[0086] In the state 614, the process updates the calculations for virtual buffer fullness, i.e., the value for d_jⁱ, d_j^p, or d_j^b. The value d_jⁱ, d_j^p, or d_j^b that is updated depends on the picture type, e.g., the d_jⁱ value is updated when an I-picture is encoded. The process updates the calculations for the virtual buffer fullness to account for the bits used to encode the macroblock. The update to the virtual buffer fullness should correspond to the technique used to allocate the bits among the macroblocks of a picture. For example, where TM5 is followed, the allocation of bits within the macroblocks of a picture can be approximately linear, i.e., constant. In one embodiment, the bits are also advantageously allocated among macroblocks based on the relative motion of a macroblock within a picture (for P-pictures and B-pictures), rather than an estimate of the relative motion.

[0087] Equations 13a, 14a, and 15a generically describe the update to the calculations for virtual buffer fullness.
d_jⁱ=d₀ⁱ+B_j−1−TMB_j−1ⁱ  (Eq. 13a)
d_j^p=d₀^p+B_j−1−TMB_j−1^p  (Eq. 14a)
d_j^b=d₀^b+B_j−1−TMB_j−1^b  (Eq. 15a)

[0088] The variable B_j corresponds to the number of bits that have already been used to encode the macroblocks in the picture that is being encoded, including the bits used in macroblock j such that the variable B_j−1 corresponds to the number of bits that have been used to encode the macroblocks up to but not including the j-th macroblock. The variables TMB_j−1ⁱ, TMB_j−1^p, and TMB_j−1^b correspond to the bits allocated to encode the macroblocks up to but not including the j-th macroblock.

[0089] Equations 13b, 14b, and 15b express calculations for virtual buffer fullness, i.e., values for d_jⁱ, d_j^p, or d_j^b, as used in the process described by TM5. Disadvantageously, the TM5 process allocates bits within a picture without regard to motion of macroblocks such that macroblocks that should have bits allocated variably to accommodate rapid motion, such as the macroblocks that encode the movement of an athlete, have the same bits allocated as macroblocks that are relatively easy to encode. $\begin{array}{cc}{d}_j^i=d_0^i+B_{j-1}-\left(\frac{T_i·\left(j-1\right)}{\mathrm{MB\_cnt}}\right)& \left(\mathrm{Eq}.13b\right)\\ d_j^p=d_0^p+B_{j-1}-\left(\frac{T_p·\left(j-1\right)}{\mathrm{MB\_cnt}}\right)& \left(\mathrm{Eq}.14b\right)\\ d_j^b=d_0^b+B_{j-1}-\left(\frac{T_b·\left(j-1\right)}{\mathrm{MB\_cnt}}\right)& \left(\mathrm{Eq}.15b\right)\end{array}$

[0090] In one embodiment, the updated values are expressed by Equations 13c, 14c, and 15c. The use of Equations 13c, 14c, and 15c permit the allocation of bits to macroblocks within a picture to be advantageously allocated based on the motion activity of a macroblock within a picture. Advantageously, such allocation can permit the bits of a picture to be allocated to macroblocks based on a computation of the relative motion of the macroblock rather than a constant amount or an estimate of the motion. The variable allocation of bits among the macroblocks of a picture will be described in greater detail later in connection with FIGS. 8A and 8B. $\begin{array}{cc}{d}_j^i=d_0^i+B_{j-1}-\left(\frac{T_i·{\mathrm{Mact\_sum}}_{j-1}}{\mathrm{MACT}}\right)& \left(\mathrm{Eq}.13c\right)\\ d_j^p=d_0^p+B_{j-1}-\left(\frac{T_p·{\mathrm{Mact\_sum}}_{j-1}}{\mathrm{MACT}}\right)& \left(\mathrm{Eq}.14c\right)\\ d_j^b=d_0^b+B_{j-1}-\left(\frac{T_b·{\mathrm{Mact\_sum}}_{j-1}}{\mathrm{MACT}}\right)& \left(\mathrm{Eq}.15c\right)\end{array}$

[0091] The variable MACT represents the sum of the motion activity of all of the macroblocks as expressed in Equation 16. The variable Mact_sum_j−1 corresponds to the sum of the motion activity of all of the macroblocks in the picture that have been encoded, i.e., the macroblocks up to but not including macroblock j, as expressed in Equation 17. $\begin{array}{cc}\mathrm{MACT}=\sum _{k=1}^{\mathrm{MB\_cnt}}{\mathrm{Mact}}_k& \left(\mathrm{Eq}.16\right)\\ {\mathrm{Mact\_sum}}_{j-1}=\sum _{k=1}^{j-1}{\mathrm{Mact}}_k& \left(\mathrm{Eq}.17\right)\end{array}$

[0092] In Equation 16, the parameter MB_cnt corresponds to the number of macroblocks in the picture and the variable Mact_k corresponds to the motion activity measure of the luminance of the k-th macroblock. A variety of techniques can be used to compute the motion activity measure such as variance computations and sum of absolute difference computations.

[0093] In another embodiment, the updated values for the occupancy of the virtual buffers d_jⁱ, d_j^p, or d_j^b are calculated based on the corresponding equations for updated virtual buffer occupancy described in Chapter 10 of the TM5 model from MPEG.

[0094] In another embodiment, the updated values for the occupancy of the virtual buffers d_jⁱ, d_j^p, or d_j^b are calculated based on Equations 13d, 14d, and 15d. $\begin{array}{cc}{d}_j^i=d_0^i+B_{j-1}-\left({\alpha }_i\frac{T_i·\left(j-1\right)}{\mathrm{MB\_cnt}}+\left(1-{\alpha }_i\right)\frac{T_i·{\mathrm{Mact\_sum}}_{j-1}}{\mathrm{MACT}}\right)& \left(\mathrm{Eq}.13d\right)\\ d_j^p=d_0^p+B_{j-1}-\left({\alpha }_p\frac{T_p·\left(j-1\right)}{\mathrm{MB\_cnt}}+\left(1-{\alpha }_p\right)\frac{T_p·{\mathrm{Mact\_sum}}_{j-1}}{\mathrm{MACT}}\right)& \left(\mathrm{Eq}.14d\right)\\ d_j^b=d_0^b+B_{j-1}-\left({\alpha }_b\frac{T_b·\left(j-1\right)}{\mathrm{MB\_cnt}}+\left(1-{\alpha }_b\right)\frac{T_b·{\mathrm{Mact\_sum}}_{j-1}}{\mathrm{MACT}}\right)& \left(\mathrm{Eq}.15d\right)\end{array}$

[0095] In Equations 13d, 14d, and 15d, α_i, α_p, and α_b correspond to weighting factors that can range from about 0 to about 1. These weighting factors α_i, α_p, and α_b permit the allocation of bits to macroblocks within a picture to be advantageously allocated based on a combination of the relatively equal proportioning from TM5 and the proportioning based on motion activity described earlier in connection with Equations 13c, 14c, and 15c. This combined allocation can advantageously compensate for bits that are relatively evenly allocated, such as bits for overhead. The values for the weighting factors α_i, α_p, and α_b can vary widely within the range of about 0 to about 1. In one embodiment, the weighting factors α_i, α_p, and α_b range from about 0 to about 0.5. For example, sample values for these weighting factors can correspond values such as 0.2, 0.3, 0.4 and 0.5. Other values within the range of about 0 to about 1 will be readily determined by one of ordinary skill in the art. One embodiment of the video encoder permits a user to configure the values for the weighting factors α_i, α_p, and α_b.

[0096] The values for the occupancy of the virtual buffers d_jⁱ, d_j^p, or d_j^b are computed for each macroblock in the picture. It will be understood, however, that the value for the first macroblock, i.e., d₁ⁱ, d₁^p, or d₁^b, is the same as the initial values set in the state 612 such that the state 614 can be skipped for the first macroblock. The process advances from the state 614 to the state 616.

[0097] In the state 616, the process computes the reference quantization parameter Q_j that is to be used to quantize macroblock j. Equation 18 expresses a computation for the reference quantization parameter Q_j. The process advances from the state 616 to a state 619. $\begin{array}{cc}{Q}_j=\left(\frac{d_j·31}{r}\right)& \left(\mathrm{Eq}.18\right)\end{array}$

[0098] In the state 619, the process computes the normalized spatial activity measures N_Sact_j for the macroblocks. In one embodiment, the process computes the normalized spatial activity measures N_Sact_j in accordance with the TM5 process and Equations 19a, 19b, 21a, 22, and 23a. Disadvantageously, the computation of the normalized spatial activity measures N_Sact_j via TM5 allocates bits to macroblocks within a picture based only on spatial activity (texture) and does not take motion into consideration. In addition, as will be explained in greater detail later in connection with Equation 23a, the TM5 process disadvantageously uses an inappropriate value in the computation of an average of the spatial activity measures Savg_act_j due to limitations in the processing sequence, which is explained in greater detail later in connection with FIGS. 8A and 8B.

[0099] In another embodiment, the process computes the normalized spatial activity measures N_Sact_j in accordance with Equations 20a, 21b, 21c, 22, and 23b. The combination of the motion activity measure used for computation of the reference quantization parameter Q_j with the modulation effect achieved through the normalized spatial activity measure advantageously permits bits to be allocated within a picture to macroblocks not only based on spatial activity (texture), but also based on motion. This can dramatically improve a picture. For example, when only spatial activity is used, areas of a picture with rapid motion, such as an area corresponding to an athlete's legs in a sporting event, are typically allocated relatively few bits, which results in visual artifacts such as a “blocky” appearance. This happens because areas of pictures with rapid motion typically exhibit relatively high spatial activity (high texture), and are then allocated relatively few bits. In addition, as will be described later in connection with Equation 23b, one embodiment further uses the actual values for spatial activity measures, which advantageously results in a better match between targeted bits and actually encoded bits, thereby decreasing the likelihood of buffer overrun or buffer underrun.

[0100] In the state 619, the activity corresponds to spatial activity within the picture to determine the texture of the picture. A variety of techniques can be used to compute the spatial activity. For example, the process can compute the spatial activity in accordance with the techniques disclosed in Chapter 10 of Test Model 5 or in accordance with new techniques that are described herein. Equation 19a illustrates a computation for the spatial activity of a macroblock j from luminance frame-organized sub-blocks and field-organized sub-blocks as set forth in Chapter 10 of Test Model 5. The intra picture spatial activity of the j-th macroblock, i.e., the texture, can be computed using Equation 19b, which corresponds to the computation that is used in TM5. $\begin{array}{cc}{\mathrm{act}}_j=1+\min \left({\mathrm{vblk}}_1,{\mathrm{vblk}}_2,\dots ,{\mathrm{vblk}}_8\right)& \left(\mathrm{Eq}.19a\right)\\ {\mathrm{vblk}}_n=\frac{1}{64}·\sum _{k-1}^{64}{\left(P_k^n-{\mathrm{P\_mean}}_n\right)}^2& \left(\mathrm{Eq}.19b\right)\end{array}$

[0101] A formula for computing the value of P_mean_n is expressed later in Equation 21a. The values for P_kⁿ correspond to the sample values from pixels in the n-th original 8 by 8 sub-block. Disadvantageously, the computation expressed in Equation 19b is relatively complicated and CPU intensive to compute, which can make real-time encoding difficult with relatively slow general purpose CPUs, such as microprocessors. Equation 19b computes the spatial activity via computation of a variance, which is referred to as L2-norm. This can be a drawback when video encoding is performed in real time and with full resolution and picture rates. As a result, real time video encoding is typically performed in conventional systems with dedicated hardware. Although dedicated hardware video encoders can process video at relatively high speeds, dedicated hardware is relatively more expensive, less supportable, and harder to upgrade than a software solution that can be executed by a general-purpose electronic device, such as a personal computer. Thus, video encoding techniques that can efficiently process video can advantageously permit a general-purpose electronic device to encode video in real time.

[0102] Equation 20a illustrates a computation for the spatial activity of macroblock j according to one embodiment. Another embodiment uses sums of absolute differences (instead of sum of squares of differences) as illustrated in Equations 19a and 19b to compute the spatial activity of macroblock j. Equation 20b illustrates a computation for the motion activity of macroblock j according to one embodiment. $\begin{array}{cc}{\mathrm{Sact}}_j=\sum _{k=1}^{256}P_k^j-{\mathrm{P\_mean}}_j& \left(\mathrm{Eq}.20a\right)\\ {\mathrm{Mact}}_j=\sum _{k=1}^{256}P_k^j-{\mathrm{P\_mean}}_j& \left(\mathrm{Eq}.20b\right)\end{array}$

[0103] In Equation 20a, the p_k^j values correspond to original luminance data. In Equation 20b, the P_k^j values correspond to either original luminance data or to motion-compensated luminance data depending on the type of macroblock. The P_kⁱ values correspond to sample values for the j-th 16 by 16 original luminance data when the macroblock is an intra macroblock. When the macroblock is an inter macroblock, the P_k^j values correspond to 16 by 16 motion compensated luminance data. A formula for computing the value of P_mean_j is expressed later in Equation 21b and 21c.

[0104] Moreover, the computations expressed in Equations 20a and 20b can advantageously permit a general-purpose electronic device to perform full picture rate and relatively high resolution video encoding using the described rate control and quantization control process in real time using software. It will be understood that the computations expressed in Equations 20a and 20b can also be used in non-real time applications and in dedicated hardware. One embodiment of a video encoding process, which was implemented in software and executed by an Intel® Pentium® 4 processor with a 3 GHz clock speed, efficiently and advantageously encoded a PAL, a SECAM, or an NTSC video data stream with a full picture rate and with full resolution (720×480 pixels) in real time.

[0105] The computations expressed in Equations 20a and 20b compute the sum of absolute differences (SAD), which is also known as an L1-norm calculation. Although the computation of the SAD can also be relatively complex, selected processors or CPUs include a specific instruction that permits the computation of the SAD in a relatively efficient manner. In one embodiment, the general-purpose electronic device corresponds to a personal computer with a CPU that is compatible with the Streaming Single Instruction/Multiple Data (SIMD) Extensions (SSE) instruction set from Intel Corporation. In another embodiment, the CPU of the general-purpose electronic device is compatible with an instruction that is the same as or is similar to the “PSADBW” instruction for packed sum of absolute differences (PSAD) of the SSE instruction set. Examples of CPUs that are compatible with some or all of the SSE instruction set include the Intel® Pentium® III processor, the Intel® Pentium® 4 processor, the Intel® Xeon™ processor, the Intel® Centrino® processor, selected versions of the Intel® Celeron® processor, selected versions of the AMD Athlon™ processor, selected versions of the AMD Duron™ processor, and the AMD Opteron™ processor. It will be understood that future CPUs that are currently in development or have yet to be developed can also be compatible with the SSE instruction set. It will also be understood that new instruction sets can be included in new processors and these new instruction sets can remain compatible with the SSE instruction set.

[0106] Equation 21 a expresses a calculation for sample values as used in Equation 19b. Equations 21b and 21c express calculations for sample values as used in Equations 20a and 20b. $\begin{array}{cc}{\mathrm{P\_mean}}_n=\frac{1}{64}·\sum _{k-1}^{64}{P}_k^n& \left(\mathrm{Eq}.21a\right)\\ {\mathrm{P\_mean}}_j=\frac{1}{256}·\sum _{k=1}^{256}{P}_k^j& \left(\mathrm{Eq}.21b\right)\\ {\mathrm{P\_mean}}_j=0& \left(\mathrm{Eq}.21c\right)\end{array}$

[0107] In one embodiment, the process performs a computation for the average of the sample values in the n-th original 8 by 8 sub-block P_mean_n according to TM5 as expressed by Equation 21a. In another embodiment, the process computes the computation for the average of sample values P_mean_j via Equations 21b and 21c. Advantageously, Equations 21b and 21c combine spatial activity (texture) computations and motion estimation computations. Equation 21b is used when the macroblock corresponds to an intra macroblock. Equation 21c is used when the macroblock corresponds to an inter macroblock.

[0108] Equation 22 expresses a computation for the normalized spatial activity measures N_Sact_j. The normalized spatial activity measures N_Sact_j are used in a state 621 to compute the quantization that is applied to the discrete cosine transform (DCT) coefficients. $\begin{array}{cc}{\mathrm{N\_Sact}}_j=\frac{\left(2·{\mathrm{Sact}}_j\right)+\mathrm{Savg\_act}}{{\mathrm{Sact}}_j+\left(2·\mathrm{Savg\_act}\right)}& \left(\mathrm{Eq}.22\right)\end{array}$

[0109] As expressed in Equation 22, the normalized spatial activity measures N_Sact_j for the j-th macroblock are computed from the spatial activity measure Sact_j for the macroblock and from an average of the spatial activity measures Savg_act. The average of the spatial activity measures Savg_act can be computed by Equation 23a or by Equation 23b. $\begin{array}{cc}\mathrm{Savg\_act}=\frac{1}{\mathrm{MB\_cnt}}·\sum _{j=1}^{\mathrm{MB\_cnt}}{\mathrm{Sact}}_j^{\mathrm{previous}}& \left(\mathrm{Eq}.23a\right)\end{array}$

[0110] The computation expressed in Equation 23a represents the computation described in TM5 and uses the spatial activity measures Sact_j from the previous picture and not from the present picture. As a result, conventional encoders that comply with TM5 compute the normalized spatial activity measures N_Sact_j expressed in Equation 22 relatively inaccurately. When a value for the average of the spatial activity measures Savg_act_j is calculated via Equation 23a, the normalized spatial activity measures N_Sact_j represents an estimate for normalization, rather than an actual calculation for normalization. The estimate provided in Equation 23a is particularly poor when the scene has changed from the previous picture to the current picture. As taught in TM5, a value of 400 can be used to initialize the average of the spatial activity measures Savg_act_j for the first picture when the average of the spatial activity measures Savg_act_j is computed from the previous picture.

[0111] Encoding via the process described in TM5 uses the previous picture for the average of the spatial activity measures Savg_act_j because the processing sequence described in TM5 processes macroblocks one-by-one as the TM5 process encodes each macroblock, such that a value for the average of the spatial activity measures Savg_act_j is not available at the time of the computation and use of the value for the normalized spatial activity measures N_Sact_j. Further details of an alternate processing sequence will be described in greater detail later in connection with FIGS. 8A and 8B. The computation expressed in Equation 23b represents an improvement over the TM5-based computation expressed in Equation 23a. $\begin{array}{cc}\mathrm{Savg\_act}=\frac{1}{\mathrm{MB\_cnt}}·\sum _{j=1}^{\mathrm{MB\_cnt}}{\mathrm{Sact}}_j^{\mathrm{current}}& \left(\mathrm{Eq}.23b\right)\end{array}$

[0112] In one embodiment, the sequence of processing of macroblocks is advantageously rearranged as will be described later in connection with FIGS. 8A and 8B. This rearrangement permits the average of the spatial activity measures Savg_act_j to be computed from the spatial activity measures Sact_j of the macroblocks in the current picture such that the value for the normalized spatial activity measures N_Sact_j is actually normalized rather than estimated. This advantageously permits the data to be relatively predictably quantized such that the amount of data used to encode a picture more accurately follows the targeted amount of data. This further advantageously reduces and/or eliminates irregularities and distortions to the values for the variables d_jⁱ, d_j^p, and d_j^b that represent the virtual buffer fullness for I-pictures, for P-pictures, and for B-pictures, respectively. In addition, it should be noted that the computation for the average of the spatial activity measures Savg_act_j expressed in Equation 23b does not need to be initialized with an arbitrary value, such as a value of 400, because the actual average is advantageously computed from the spatial activity measures Sact_j of the picture that is currently being encoded. The process advances from the state 619 to the state 621. Advantageously, this permits calculation of actual motion activity measures, needed for the calculation of virtual buffer fullness status, as shown in Equations 13-17.

[0113] In the state 621, the process computes the quantization parameter mquant_j. The quantization parameter mquant_j is used to quantize the encoded macroblock j. It will be understood that the quantization parameter mquant_j can be used in the state 621 or can be stored and used later. Equation 23 expresses a computation for the quantization parameter mquant_j.
mquant_j=Q_j·N_—Sact_j  (Eq. 23)

[0114] In Equation 23, Q_j corresponds to the reference quantization parameter described earlier in connection with Equation 18 and N_act_j corresponds to the normalized spatial activity measures N_Sact_j described earlier in connection with Equation 22. In one embodiment, the process further inspects the computed quantization parameter mquant_j and limits its value to prevent undesirable clipping of a resulting quantized level QAC(ij). For example, where one embodiment of the process is used to encode video according to the MPEG-1 standard, the process detects that the calculated value for the quantization parameter mquant_j corresponds to 2, and automatically substitutes a value of 4. The quantization parameter mquant_j is later used in the macroblock encoding process to generate values for the quantized level QAC(ij). However, in MPEG-1, a value for the quantized level QAC(ij) is clipped to the range between −255 and 255 to fit within 8 bits. This clipping of data can result in visible artifacts, which can advantageously be avoided by limiting the value of a quantization parameter mquant_j to a value that prevents the clipping of the resulting quantized level, thereby advantageously improving picture quality.

[0115] In one embodiment, the process can further reset values for occupancy of virtual buffers (d_jⁱ, d_j^p, and d_j^b) and for the quantization parameter mquant_j in response to selected stimuli as will be described in greater detail later in connection with FIG. 9A. The process advances from the state 621 to a state 623.

[0116] In the state 623, the process encodes the j-th macroblock. The process encodes the j-th macroblock using the quantization parameter mquant_j computed earlier in the state 616. The encoding techniques can include, for example, the computation of discrete cosine transforms, motion vectors, and the like. In one embodiment, the process can selectively skip the encoding of macroblocks in B-pictures as will be described in greater detail later in connection with FIG. 11. The process advances from advances from the state 623 to a decision block 625.

[0117] In the decision block 625, the process determines whether all the macroblocks in the picture have been processed by encoding in the state 616 or by skipping as will be described in connection with FIG. 11. The process proceeds from the decision block 625 to a state 627 when the process has completed the encoding or skipping processing of the macroblocks in the picture. Otherwise, the process returns from the decision block 625 to the state 614 to continue to process the next macroblock.

[0118] In the state 627, the process stores the final occupancy value of the virtual buffers as an initial condition for encoding of the next picture of the same type. For example, the final occupancy value for the relevant virtual buffer of the present frame, i.e., the value for d_jⁱ, d_j^p, or d_j^b, when j is equal to MB_cnt, is saved so that it can be used as a starting value for d₀ⁱ, d₀^p, or d₀^b, respectively, for the next picture of the same type. In some circumstances, the number of bits used for encoding can be relatively low for a sustained period of time so that bit or byte stuffing is used to increase the number of bits used in encoding. This prevents a buffer overrun condition in the decoder buffer. However, the use of bit stuffing can undesirably distort the occupancy value in the corresponding virtual buffer, which can then result in instability in the encoder. In one embodiment, the rate control and quantization control process includes one or more techniques that advantageously ameliorate against the effects of bit stuffing. Examples of such techniques will be described in greater detail later in connection with FIGS. 9A and 9B. The process advances from the state 627 to a decision block 630.

[0119] In the decision block 630, the illustrated process has completed the processing for the picture and determines whether the picture that was processed corresponds to the last picture in the group of pictures (GOP). This can be accomplished by monitoring the values remaining in the number of P-pictures N_p and the number of B-pictures N_b described earlier in connection with the state 606. The process proceeds from the decision block 630 to a state 632 when there are pictures that remain to be processed in the group of pictures. Otherwise, i.e., when the process has completed processing of the group of pictures, the process proceeds from the decision block 630 to a decision block 634.

[0120] In the state 632, the process updates the appropriate value in the number of P-pictures N_p or the number of B-pictures N_b and advances to a state 636 to initiate the processing of the next picture in the group of pictures. It will be understood that the next picture to be processed may not be the next picture to be displayed because of possible reordering of pictures during encoding.

[0121] In the state 636, the process updates the corresponding complexity estimators X_i, X_p, and X_b based on the picture that just been encoded. For example, if an I-picture had just been encoded, the process updates the complexity estimator X_i for the I-pictures as expressed in Equation 24. If the picture that had just been encoded was a P-picture or was a B-picture, the process updates the corresponding complexity estimator X_p or X_b, respectively, as expressed in Equation 25 and in Equation 26.
X_i=S_iQ_i  (Eq. 24)
X_p=S_pQ_p  (Eq. 25)
X_b=S_bQ_b  (Eq. 26)

[0122] In Equations 24, 25, and 26, the value of S_i, S_p, or S_b corresponds to the number of bits generated or used to encode the picture for a picture of type I-picture, P-picture, or B-picture, respectively. The value of Q_i, Q_p, and Q_b corresponds to the average of the values for the quantization parameter mquant_j that were used to quantize the macroblocks in the picture. The process advances from the state 636 to a state 638.

[0123] In the state 638, the process updates the remaining number of bits R allocated to the group of pictures. The update to the remaining number of bits R allocated to the group of pictures depends on whether the next picture to be encoded is a picture from the existing group of pictures or whether the next picture to be encoded is the first picture in a new group of pictures. Both Equations 27 and 28 are used when the next picture to be processed is the first picture in a new group of pictures. When the next picture to be processed is another picture in the same group of pictures as the previously processed picture, then only Equation 27 is used. It will be understood that Equations 27 and 28 represent assignment statements for the value of R, such that a new value for R is represented to the left of the “=” sign and a previous value for R is represented to the right of the “=” sign.
R=R−S_(i,p,b)  (Eq. 27)
R=G+R  (Eq. 28)

[0124] In Equation 27, the process computes the new value for the remaining number of bits R allocated to the group of pictures by taking the previous value for R and subtracting the number of bits S_{(i p,b)} that had been used to encode the picture that had just been encoded. The number of bits S_{(i p,b)} that had been used to encode the picture is also used to calculate the VBV buffer model occupancy as will be described in greater detail later in connection with FIG. 7. The computation expressed in Equation 27 is performed for each picture after it has been encoded. When the picture that has just been encoded is the last picture in a group of pictures such that the next picture to be encoded is the first picture in a new group of pictures, the computation expressed in Equation 27 is further nested with the computation expressed in Equation 28. In Equation 28, the process adds to a remaining amount in R, which can be positive or negative, a value of G. The variable G was described earlier in connection with Equation 5. The value of G is based on the new group of pictures to be encoded and corresponds to the number of bits that can be transferred by the data channel in the amount of time corresponding to the length of the presentation time for the new group of pictures. The process returns from the state 638 to the state 610 to continue to the video encoding process as described earlier.

[0125] Returning now to the decision block 634, at this point in the process, the process has completed the encoding of a picture that was the last picture in a group of pictures. In the decision block 634, the process determines whether it has completed with the encoding of the video sequence. It will be understood that the process can be used to encode video of practically indefinite duration, such as broadcast video, and can continue to encode video endlessly. The process proceeds from the decision block 634 to a state 640 when there is another group of pictures to be processed. Otherwise, the process ends.

[0126] In the state 640, the process receives the next group of pictures. It will be understood that in another embodiment, the process may retrieve only a portion of the next group of pictures in the state 640 and retrieve remaining portions later. In one embodiment, the state 640 is relatively similar to the state 602. The process advances from the state 640 to a state 642.

[0127] In the state 642, the process receives the mode or type of encoding that is to be applied to the pictures in the group of pictures. In the illustrated rate control and quantization control process, the decision as to which mode or type of encoding is to be used for each picture in the group of pictures is made before the pictures are processed by the rate control and quantization control process. In one embodiment, the state 642 is relatively similar to the state 604. The process advances from the state 642 to a state 644.

[0128] In the state 644, the process determines the number of P-pictures N_p and the number of B-pictures N_b in the next group of pictures to be encoded. In one embodiment, the state 644 is relatively similar to the state 606. The process advances from the state 644 to the state 636, which was described in greater detail earlier, to continue with the encoding process.

Control With VBV Buffer Model Occupancy Levels

[0129] FIG. 7 is a flowchart that generally illustrates a process for adjusting a targeted bit allocation based on an occupancy level of a virtual buffer. To illustrate the operation of the process, the process will be described in connection with MPEG-1 and MPEG-2 video encoding so that the virtual buffer corresponds to the video buffer verifier (VBV) buffer model. The VBV buffer model is a conceptual model that is used by the encoder to model the buffer occupancy levels in a decoder. It will be apparent to one of ordinary skill in the art that other buffer models can be used with other video encoding standards. Monitoring of VBV buffer model levels will be described now in greater detail before further discussion of FIG. 7.

[0130] As described earlier in connection with FIG. 4, the VBV buffer model anticipates or predicts buffer levels in the decoder buffer. The occupancy level of the decoder buffer is approximately inverse to the occupancy level of the encoder buffer, such that a relatively high occupancy level in the VBV buffer model indicates that relatively few bits are being used to encode the video sequence, and a relatively low occupancy level in the VBV buffer model indicates that relatively many bits are being used to encode the video sequence.

[0131] The occupancy level V_status of the VBV buffer model is computed and monitored. In one embodiment, the occupancy level V_status of the VBV buffer model is compared to a predetermined threshold, and the encoding can be adapted in response to the comparison as will be described in greater detail later in connection with FIG. 11. In another embodiment, the occupancy level V_status of the VBV buffer model is used to adaptively adjust a target number of bits T_i, T_p, or T_b for a picture to be encoded. A computation for the occupancy level V_status is expressed in Equation 29. $\begin{array}{cc}{V}_{\mathrm{status}}=V_{\mathrm{status}}-S_{\left(i,p,b\right)}+\frac{\mathrm{bit\_rate}}{\mathrm{picture\_rate}}& \left(\mathrm{Eq}.29\right)\end{array}$

[0132] Equation 29 represents an assignment statement for the value of the occupancy level V_status. A new value for the occupancy level V_status is represented at the left of the “=” sign, and a previous value for the occupancy level V_status is represented to the right of the “=” sign. In one embodiment, the value of the occupancy level V_status is initialized to a target value for the VBV buffer model. An example of a target value is ⅞'s of the full capacity of the VBV buffer model. In another embodiment, the value of V_status is initialized to a buffer occupancy that corresponds to a specified VBV-delay value. Other initialization values can be readily determined by one of ordinary skill in the art.

[0133] In Equation 29, the occupancy of the VBV buffer model is computed as follows. The number of bits S_(i,p,b) that had been used to encode the picture just encoded is subtracted from the previous value for the occupancy level V_status and the number of bits that would be transmitted in the time period corresponding to a “frame” or picture is added to the value for the occupancy level V_status. As illustrated in Equation 29, the number of bits that would be transmitted in the frame is equal to bit rate times the inverse of the frame rate. The computation expressed in Equation 29 is adapted to update the occupancy level V_status for each picture processed. In another embodiment, the expression is modified to update the occupancy level V_status for less than each picture, such as every other picture.

[0134] As will be described later in connection with FIG. 7, one embodiment of the process compares the target number of bits for a picture T_i, T_p, or T_b to a threshold T_mid, and adjusts the target number of bits T_i, T_p, or T_b in response to the comparison. This advantageously assists the video encoder to produce a data stream that is compliant with VBV to protect against buffer underrun or buffer overrun in the decoder.

[0135] One embodiment uses five parameters related to VBV buffer model occupancy levels for control. It will be understood that in other embodiments, fewer than five parameters or more than five parameters can also be used. The parameters can vary in a very broad range and can include fixed parameters, variable parameters, adaptable parameters, user-customizable parameters, and the like. In one embodiment, the following parameters are used (in decreasing order of occupancy): V_high, V_target, V_mid, V_low, and V_critical.

[0136] V_high corresponds to a relatively high value for the occupancy of the VBV buffer model. In one embodiment, the process strives to control encoding such that the occupancy of the VBV buffer model is maintained below V_high.

[0137] V_target corresponds to an occupancy level for the VBV buffer model that is desired. In one embodiment, the desired buffer occupancy level V_target can be configured by a user.

[0138] V_mid corresponds to an occupancy level that is about half of the capacity of the VBV buffer model.

[0139] V_low corresponds to a relatively low value for the occupancy of the VBV buffer model. In one embodiment, the process strives to control encoding such that the occupancy of the VBV buffer model is maintained above V_low.

[0140] V_critical corresponds to an even lower occupancy level than V_low. In one embodiment, when the occupancy of the VBV buffer model falls below V_critical, the process proceeds to skip macroblocks in B-pictures as will be described in greater detail later in connection with FIG. 11.

[0141] Table II illustrates sample values for threshold levels. Other suitable values will be readily determined by one of ordinary skill in the art.

[0142]

	TABLE II
	Threshold	Sample Value
	Vhigh	about 63/64 of VBV buffer model size
	Vtarget	about ⅞ of VBV buffer model size
	Vmid	about ½ of VBV buffer model size
	Vlow	about ⅜ of VBV buffer model size
	Vcritical	about ¼ of VBV buffer model size

[0143] The sample values listed in Table II are advantageously scaled to the VBV buffer model size. As described in greater detail earlier in connection with FIG. 4, the VBV buffer model size is approximately 224 kB for MPEG-2 and is approximately 40 kB for MPEG-1. It will be understood by one of ordinary skill in the art that the size of a virtual buffer model, such as the VBV buffer model for MPEG-1 and MPEG-2, can vary according with the video encoding standard used and the application scenario.

[0144] Returning now to FIG. 7, the process illustrated in FIG. 7 adjusts a targeted bit allocation T_i, T_p, or T_b for a picture based at least in part on the occupancy level V_status of the VBV buffer model. In one embodiment, the process illustrated in FIG. 7 is incorporated in the state 610 of the process illustrated in FIG. 6. The process can start at an optional decision block 710, where the process compares the value of the targeted bit allocation T_i, T_p, or T_b (generically written as T_(i,p,b) in FIG. 7) to one or more target thresholds, such as to T_mid or to T_high. For example, the target threshold T_mid can be selected such that the adjustment process is invoked when the VBV buffer model occupancy level is relatively low. In another example, the target threshold T_high can be selected such that the adjustment process is invoked when the VBV buffer model occupancy is relatively high. In one embodiment, only one of the target thresholds T_mid or T_high is used, in another embodiment, both target thresholds are used, and in yet another embodiment, the optional decision block 710 is not present and neither target threshold is used. In the illustrated embodiment, the adjustment process is invoked in response to the VBV buffer model occupancy level and to the number of bits allocated to the picture to be encoded. The computation of the targeted bit allocation T_i, T_p, or T_b can be performed as described earlier in connection with the state 610 and Equations 6, 7, and 8 of FIG. 6. Equation 30a expresses a sample computation for the target threshold T_mid. Equation 30b expresses a sample computation for the target threshold T_high. $\begin{array}{cc}{T}_{\mathrm{mid}}=V_{\mathrm{status}}-V_{\mathrm{mid}}& \left(\mathrm{Eq}.30a\right)\\ T_{\mathrm{high}}=V_{\mathrm{status}}-V_{\mathrm{high}}+\frac{\mathrm{bit\_rate}}{\mathrm{picture\_rate}}& \left(\mathrm{Eq}.30b\right)\end{array}$

[0145] The illustrated embodiment of the process proceeds from the optional decision block 710 to a state 720 when the targeted bit allocation T_i, T_p, or T_b exceeds the target threshold T_mid or when the targeted bit allocation T_i, T_p, or T_b is less than the target threshold T_high. It will be understood that in another embodiment or configuration, where the optional decision block 710 is not present, the process can start at the state 720. When the targeted bit allocation T_i, T_p, or T_b exceeds the target threshold T_mid, the VBV buffer model occupancy is relatively low. In the illustrated embodiment, the target threshold T_mid, is selected such that the adjustment to the targeted bit allocation occurs when a picture is allocated enough bits such that, without adjustment, the VBV buffer model occupancy would fall or would stay below V_mid. Other thresholds will be readily determined by one of ordinary skill in the art.

[0146] When the targeted bit allocation T_i, T_p, or T_b does not exceed the target threshold T_mid and the targeted bit allocation T_i, T_p, or T_b is not less than the target threshold T_high, the illustrated process proceeds from the optional decision block 710 to a decision block 730. It will be understood that where the optional decision block 710 is not present or is not used, the process can begin at the state 720, which then proceeds to the decision block 730. In another embodiment, when the targeted bit allocation T_i, T_p, or T_b does not exceed the target threshold T_mid and the targeted bit allocation T_i, T_p, or T_b is not less than the target threshold T_high, the process proceeds to end from the optional decision block 710, such as, for example, by proceeding to the state 612 of the process described in connection with FIG. 6. In the illustrated optional decision block 710, the comparison uses the same target thresholds T_mid and/or T_high for I-pictures, for P-pictures, and for B-pictures. In another embodiment, the target thresholds T_mid and/or T_high varies depending on the picture type.

[0147] In the state 720, which is entered when the targeted bit allocation T_i, T_p, or T_b exceeds the target threshold T_mid, or when the targeted bit allocation T_i, T_p, or T_b is less than the target threshold Thigh, the process adjusts the value of the targeted bit allocation T_i, T_p, or T_b to reduce the number of bits allocated to the picture. In another embodiment, the process starts at the state 720. For example, one embodiment of the process is configurable by a user such that the process does not have the optional decision block 710 and instead, starts at the state 720. For example, the adjustment to the T_i, T_p, or T_b can be configured to decrease the number of bits. Advantageously, when fewer bits are used to encode a picture, the VBV buffer model occupancy level, and correspondingly, a decoder's buffer occupancy level, can increase. Equation 31 illustrates a general formula for the adjustment.
T_(i,p,b)=α·T_(i,p,b)  (Eq. 31)

[0148] In Equation 31, the adjustment factor a can be less than unity such that the targeted bit allocation T_i, T_p, or T_b after adjustment is smaller than originally calculated. In one embodiment, the adjustment factor a can also correspond to values greater than unity such that the targeted bit allocation T_i, T_p, or T_b after adjustment is larger than originally calculated. For clarity, the adjustment of Equation 31 illustrates an adjustment to a separately calculated targeted bit allocation T_i, T_p, or T_b. However, it will be understood that the adjustment can also be incorporated in the initial calculation of the targeted bit allocation T_i, T_p, or T_b. It will be understood that Equation 31 corresponds to an assignment statement such that the value to the right of the “=” corresponds to the targeted bit allocation T_i, T_p, or T_b before adjustment, and the value to the left of the “=” corresponds to the targeted bit allocation T_i, T_p, or T_b after adjustment. Equation 32 expresses a sample computation for the adjustment factor α. $\begin{array}{cc}\alpha =1+\frac{V_{\mathrm{status}}-V_{\mathrm{target}}}{V_{\mathrm{high}}-V_{\mathrm{low}}}& \left(\mathrm{Eq}.32\right)\end{array}$

[0149] As illustrated in Equation 32, the adjustment factor α is less than unity when V_status is less than V_target, and the adjustment factor α is greater than unity when V_status is greater than V_target. A net effect of the adjustment expressed in Equation 31 is to trend the occupancy level of the VBV buffer model to the desired occupancy level V_target.

[0150] It should be noted that when the targeted bit allocation T_i, T_p, or T_b exceeds the target threshold T_mid in the optional decision block 710, the value for the VBV buffer model occupancy V_status will typically be less than the value for the desired VBV occupancy level V_target such that adjustment factor a is less than unity. Advantageously, the targeted bit allocation can be reduced by an amount related to how much below the VBV buffer model occupancy V_status is from the desired VBV occupancy level V_target. When the targeted bit allocation T_i, T_p, or T_b is less than the target threshold T_high, the value for the VBV buffer model occupancy V_status will typically be higher than the value for the desired VBV occupancy level V_target such that adjustment factor α is greater than unity. Advantageously, the targeted bit allocation can be increased by an amount related to how much above the VBV buffer model occupancy V_status is from the desired VBV occupancy level V_target. The process advances from the state 720 to the decision block 730.

[0151] In the decision block 730, the process determines whether the targeted bit allocation T_i, T_p, or T_b, with or without adjustment by the state 720, falls within specified limits. These limits can advantageously be used to prevent a value for the targeted bit allocation T_i, T_p, or T_b from resulting in buffer underrun or buffer overrun. These limits can be predetermined or can advantageously be adapted to the targeted bit allocation T_i, T_p, or T_b and the VBV buffer model occupancy level V_status. When the targeted bit allocation T_i, T_p, or T_b falls outside the limits, the process proceeds from the decision block 730 to a state 740 to bind the targeted bit allocation T_i, T_p, or T_b to the limits. Otherwise, the process ends without further adjustment to the targeted bit allocation T_i, T_p, or T_b.

[0152] Equation 33 illustrates a sample computation for an upper limit T_max for the targeted bit allocation T_i, T_p, or T_b. Equation 34 illustrates a sample computation for a lower limit T_min for the targeted bit allocation T_i, T_p, or T_b. $\begin{array}{cc}{T}_{\max }=V_{\mathrm{status}}-V_{\mathrm{low}}& \left(\mathrm{Eq}.33\right)\\ T_{\min }=\max \left(V_{\mathrm{status}}+\frac{\mathrm{bit\_rate}}{\mathrm{picture\_rate}}-V_{\mathrm{high}},0\right)& \left(\mathrm{Eq}.34\right)\end{array}$

[0153] It will be understood that when the targeted bit allocation T_i, T_p, or T_b exceeds the upper limit T_max, the targeted bit allocation T_i, T_p, or T_b is reassigned the value of the upper limit T_max, and when the targeted bit allocation T_i, T_p, or T_b is below the lower limit T_min, the targeted bit allocation T_i, T_p, or T_b is reassigned the value of the lower limit T_min.

[0154] The application of the upper limit T_max expressed in Equation 33 advantageously limits a relatively high value for the targeted bit allocation T_i, T_p, or T_b such that the VBV buffer model occupancy level stays above the lower desired occupancy limit level V_low for the VBV buffer model. The application of the lower limit T_min expressed in Equation 34 advantageously limits a relatively low value for the targeted bit allocation T_i, T_p, or T_b such that the buffer occupancy level stays below the upper desired occupancy limit level V_high, even after the accumulating data over time at the constant bit rate of the data channel. The lower limit T_min corresponds to the higher of the quantities separated by the comma in the expression. Other values for the upper limit T_max and for the lower limit T_min will be readily determined by one of ordinary skill in the art. It will be understood that the targeted bit allocation T_i, T_p, or T_b represents a target for the encoder to achieve and that there may be relatively small variances from the target and the number of bits actually used to encode a picture such that the buffer occupancy level V_status may still deviate slightly from the desired occupancy limit levels V_low and V_high.

[0155] After processing in the state 740, the adjustment process ends. For example, where the adjustment process depicted in FIG. 7 is incorporated in the state 610 of the rate control and quantization control process illustrated in FIG. 6, the process can continue processing from the state 610.

[0156] It will be appreciated by the skilled practitioner that the illustrated process can be modified in a variety of ways without departing from the spirit and scope of the invention. For example, in another embodiment, various portions of the illustrated process can be combined, can be rearranged in an alternate sequence, can be removed, and the like. For example, in one embodiment, the optional decision block 710 is not present. In another embodiment, the decision block 730 and the state 740 are optional and need not be present.

Macroblock Processing Sequence

[0157] FIG. 8A is a flowchart that generally illustrates a sequence of processing macroblocks according to the prior art. FIG. 8B is a flowchart that generally illustrates a sequence of processing macroblocks according to one embodiment. The processing sequence illustrated in FIG. 8B advantageously permits the spatial activity and/or motion activity for the macroblocks of a picture to be calculated such that actual values can be used in computations of sums and averages as opposed to estimates of sums and averages from computations of a prior picture.

[0158] The conventional sequence depicted in FIG. 8A starts at a state 802. In the state 802, the process performs a computation for spatial activity (texture) and/or for motion estimation for a single macroblock. The process advances from the state 802 to a state 804.

[0159] In the state 804, the process uses the computation of spatial activity and/or motion estimation to perform a discrete cosine transformation (DCT) of the macroblock. The computation of spatial activity is typically normalized with a total value of spatial activity. However, at this point in the process, the computations for spatial activity have not been completed for the picture that is being encoded. As a result, an estimate from a previous picture is used. For example, the total spatial activity from the prior picture is borrowed to compute an average. In another example, motion estimation from a previous picture can also be borrowed. Whether or not these estimates are close to the actual values is a matter of chance. When there is a scene change between the prior picture and the picture that is being encoded, the estimates can be quite inaccurate. These inaccuracies can impair picture quality and lead to mismatches between the number of bits targeted for encoding of the picture and the number of bits actually used to encode the picture. These variances in the number of bits consumed to encode a picture can disadvantageously lead to buffer underrun or to buffer overrun. The process advances from the state 804 to a state 806.

[0160] In the state 806, the process performs variable length coding (VLC) for the DCT coefficients of the macroblock. The VLC compresses the DCT coefficients. The process advances from the state 806 to a decision block 808.

[0161] In the decision block 808, the process determines whether it has completed encoding all the macroblocks in the picture. The process returns from the decision block 808 to the state 802 when there are macroblocks remaining to be encoded. Otherwise, the process proceeds to end until restarted.

[0162] A rearranged sequence according to one embodiment is depicted in FIG. 8b and starts at a state 852. In the state 852, the process performs computations for spatial activity and/or motion estimation for all the macroblocks in the picture that is being encoded. This advantageously permits sums and averages of the spatial activities and/or motion estimates to be advantageously computed with actual numbers and not with estimates, and is further advantageously accurate even with a scene change before the picture that is presently encoded. In another example of advantages, in TM5, an average of the spatial activity measures Savg_act_j of 400 is used for the first picture as a “guess” of the measure. By processing the spatial activity of all the macroblocks before the spatial activities are used, the average of the spatial activity measures Savg_act_j can be directly computed and a speculative “guess” can advantageously be avoided.

[0163] Further advantageously, the use of actual sums and averages permits the actual number of bits used to encode a picture to match with the targeted bit allocation with relatively higher accuracy. This advantageously decreases the chances of undesirable buffer underrun or buffer overrun and can increase picture quality. In one embodiment, the actual motion estimation for a macroblock is used to allocate bits among the macroblocks such that macroblocks with relatively high motion are allocated a relatively high number of bits. By contrast, in a conventional system with macroblock by macroblock processing, the bits for macroblocks are typically allocated among macroblocks by the relative motion of the macroblock in a prior picture, which may or may not be accurate. The process advances from the state 852 to a state 854.

[0164] In the state 854, the process performs the DCT computations for all of the macroblocks in the picture. The process advances from the state 854 to a state 856.

[0165] In the state 856, the process performs VLC for the DCT coefficients of all of the macroblocks in the picture. The process then ends until restarted.

[0166] In another embodiment, the process performs the computation of spatial activity and/or motion estimation for all the macroblocks as described in connection with the state 852, but then loops repetitively around a state to perform DCT computations and another state to perform VLC for macroblocks until processing of the macroblocks of the picture is complete.

Bit Stuffing

[0167] Bit stuffing or byte stuffing is a technique that is commonly used by an encoder to protect against generating a data stream that would otherwise lead to a decoder buffer overrun. When the number of bits that is used to encode a picture is relatively low for a sustained period of time, the decoder retrieves data from the decoder buffer at a slower rate than the rate at which the data channel adds data to the decoder buffer. When this accumulation of data continues for a sustained period of time such that the decoder buffer fills to capacity, data carried by the data channel can be lost. An example of a sequence of pictures that can be relatively highly compressed such that bit stuffing may be invoked is a sequence of pictures, where each picture is virtually completely black. To address this disparity in data rates such that buffer overrun does not occur, the encoder embeds data in the data stream that is not used, but consumes space. This process is known as bit stuffing.

[0168] Bit stuffing can be implemented in a variety of places in an encoding process. In one embodiment, bit stuffing is implemented when appropriate after the state 632 and before the state 636 in the encoding process described in connection with FIG. 6. In one embodiment, the encoding process invokes bit stuffing when the occupancy of the VBV buffer model attains a predetermined level, such as the V_high level described earlier in connection with FIG. 7. In one embodiment, bit stuffing is invoked when the VBV buffer model occupancy is about 63/64 of the capacity of the VBV buffer model.

[0169] Though beneficial to resolving decoder buffer overrun problems, bit stuffing can introduce other problems to the encoding process. The inclusion of bits used in bit stuffing can also be an undesirable solution. The addition of bits used in bit stuffing in a computation for the number of bits used to encode a picture S_(i,p,b) can indicate to the encoder that more bits are being used to encode the pictures than were initially targeted. This can further be interpreted as an indication to encode pictures with reduced quality to decrease the number of bits used to encode pictures. Over a period of time, this can lead to an even further decrease in the number of bits used to encode the pictures, with proportionally even more bits used in bit stuffing. With relatively many bits used in bit stuffing, relatively few bits remain to actually encode the pictures, which then reduces the quality of the encoded pictures over time.

[0170] FIG. 9A illustrates a process that advantageously stabilizes the encoding process, thereby reducing or eliminating the tendency for bit stuffing to destabilize an encoding process and the tendency for the picture quality to degrade over time. As will be described later, the process depicted in FIG. 9A can be implemented in a variety of locations within an encoding process.

[0171] It will be appreciated by the skilled practitioner that the illustrated process can be modified in a variety of ways without departing from the spirit and scope of the invention. For example, in another embodiment, various portions of the illustrated process can be combined, can be rearranged in an alternate sequence, can be removed, and the like. The process can begin at a decision block 902 or at a decision block 904. In one embodiment, only one of the decision block 902 or the decision block 904 is present in the process. In the illustrated embodiment, both the decision block 902 and the decision block 904 are present in the process. For example, the process can start at the decision block 902 prior to the encoding of a picture, and the process can start at the decision block 904 after the encoding of a picture. For example, the start of process of FIG. 9A at the decision block 902 can be incorporated after the state 612 and before the state 614 of the rate control and quantization control process described in connection with FIG. 6. In another example, the start of the process of FIG. 9A at the decision block 904 can be incorporated at the state 627 of the process of FIG. 6.

[0172] In the decision block 902, the process determines whether there has been a scene change between the picture that is being encoded and the previous picture encoded. The determination of a scene change can be performed prior to the encoding of a picture. In one embodiment, the decision block 902 is optional. A variety of methods can be used to determine whether there has been a scene change. In one embodiment, the process reuses the results of a computation that is used to encode the picture, such as the results of a sum of absolute differences (SAD) measurement. In one embodiment, scene change detection varies according to the picture type. In one embodiment, for I-pictures, the average spatial activity Sact_avg for the current picture is compared to the corresponding previous average spatial activity. For example, when the current activity is at least 2 times or less than half that of the previous I-picture, a scene change is detected. Other factors that can be used, such as 3 times and ⅓, 4 times and ¼ or a combination of these will be readily determined by one of ordinary skill in the art. In addition, one embodiment imposes an additional criterion for a minimum number of pictures to pass since the previous scene change has been declared in order to declare a new scene change. For P-pictures, the average of motion activity can be used instead of the average spatial activity to detect a scene change, together with a relative comparison factor such as (2, ½), (3, ⅓), (4, ¼) and the like. To increase the robustness of the decision, one embodiment further uses a minimum average motion activity measure for the current P picture, since average motion activity by itself can indicate relatively high motion, which can be attributed to a scene change. For example, values of minimum average motion activity measure in the range of about 1000 to about 4000 can be used to indicate relatively high motion

[0173] The process proceeds from the decision block 902 to end such as, for example, by entering the state 614 when the process determines that there has been no scene change. In addition, it will be understood that there may be other portions of the encoding process which determine whether there has been a scene change, and where applicable, a previous determination can be reused in the decision block 902 by inspection of the state of a flag or semaphore indicating whether there has been a scene change. When the process determines that there has been a scene change, the process proceeds from the decision block to a sub-process 906.

[0174] In the decision block 904, the process determines whether the encoding process is in a critical state. In an alternate embodiment of the process, only one of the decision block 902 or the decision block 904 is present, and the other is optional. Where the decision block 904 is present in the process, the monitoring of the occupancy of the VBV buffer model can be invoked after the encoding of a picture. The criteria for determining that the encoding process is in a critical state can vary in a very broad range. In one embodiment, the critical state corresponds to when bit stuffing is performed by the encoding process when a value for the quantization parameter mquant_j is not relatively low, such as not at its lowest possible value. The value for the quantization parameter mquant_j that will correspond to relatively low values, such as the lowest possible value, will vary according to the syntax of the encoding standard. The process proceeds from the decision block 904 to the sub-process 906 when the occupancy of the VBV buffer model is determined to be in the critical state. Otherwise, the process proceeds to end such as, for example, by entering the state 627 of the process described in connection with FIG. 6.

[0175] In the sub-process 906, the process normalizes the virtual buffer occupancy values for the initial conditions as represented by the variables d₀ⁱ, d₀^p, and d₀^b described earlier in connection with the state 612. The normalized values can be computed by a variety of techniques. In the illustrated sub-process 906, the normalized values depend on the occupancy level of the VBV buffer model. The illustrated sub-process 906 includes a state 908, a decision block 910, a state 912, and a state 914.

[0176] In the state 908, one embodiment of the process calculates values for a sum and a delta as set forth in Equations 35 and 36a or 36b.
sum=d₀ⁱ+d₀^p+d₀^b  (Eq. 35)
delta=vbv_buffer_size−V_status  (Eq. 36a)
delta=V_initial−V_status  (Eq. 36b)

[0177] For Equation 35, the values for the virtual buffer occupancy levels for the initial conditions can be obtained by application of Equations 9, 10, and 11 as described in greater detail earlier in connection with the state 612 of FIG. 6. As illustrated in Equations 36a and 36b, delta increases with a decreasing occupancy level in a buffer model. In Equation 36a, the variable vbv_buffer_size relates to the capacity of the VBV buffer model that is used for encoding. In Equation 36b, the variable V_initial relates to an initialization value for the occupancy level of the VBV buffer model. In one embodiment, the value of V_initial is about ⅞'s of the capacity of the VBV buffer model. In another embodiment, instead of V_initial, the process can use a target occupancy level such as V_target, but it should be noted that the initialization value and the target occupancy can be the same value. In another embodiment, delta can be based on a different quantity related to the size of the buffer model subtracted by the occupancy level of the buffer model. The size or capacity of the VBV buffer model can vary according to the standard that is used for encoding. For example, as described earlier in connection with FIG. 4, the MPEG-1 and the MPEG-2 encoding standards specify a VBV buffer size or about 40 kB and about 224 kB, respectively. Other standards can specify amounts of memory capacity for the VBV buffer model. The process advances from the state 908 to the decision block 910.

[0178] In the decision block 910, the process determines whether the value for sum is less than the value for a predetermined threshold T_norm. The value of the predetermined threshold T_norm should correspond to some value that indicates a usable range. For example, one such value for the predetermined threshold T_norm is zero. Other values will be readily determined by one of ordinary skill in the art. The process proceeds from the decision block 910 to the state 912 when the value for sum is less than the value T_norm. Otherwise, the process proceeds from the decision block 910 to the state 914.

[0179] The value for delta corresponds to the unoccupied space in the VBV buffer model for Equation 36a or to the discrepancy between the initial VBV buffer model status and the current VBV buffer model status in Equation 36b. It will be understood that other comparisons can be made between the sum of the virtual buffer levels and the unoccupied levels. For example, in another embodiment, a less than or equal to comparison can be made, an offset can be included, etc.

[0180] In the state 912, one embodiment of the process reassigns the virtual buffer occupancy values for the initial conditions d₀ⁱ, d₀^p, and d₀^b with normalized values according to Equations 37, 38, and 39.
d₀ⁱ=delta·frⁱ  (Eq. 37)
d₀^p=delta·fr^p  (Eq. 38)
d₀^b=delta·fr^b  (Eq. 39)

[0181] In Equations 37, 38, and 39, the value for delta can be calculated from Equation 36, and the values for frⁱ, fr^p, and fr^b can vary in a very broad range. The values for frⁱ, fr^p, and fr^b will typically range between 0 and 1 and can be the same value or different values. Further, in one embodiment, the values for frⁱ, fr^p, and fr^b are selected such that they sum to a value of approximately 1, such as the value of 1. In one embodiment, the values for frⁱ, fr^p, and fr^b correspond to about 5/17, about 5/17, and about 7/17, respectively. Other values for frⁱ, fr^p, and fr^b will be readily determined by one of ordinary skill in the art. The process can then end by, for example, entering the state 614 of the process described in connection with FIG. 6.

[0182] Returning to the state 914, at this point in the process, the process has determined that the value for sum is not less than the value for T_norm. In the state 914, one embodiment of the process reassigns the values of the virtual buffer occupancy variables for the initial conditions d₀ⁱ, d₀^p, and d₀^b with normalized values according to Equations 40, 41, and 42. $\begin{array}{cc}{d}_0^i=d_0^i·\frac{\mathrm{delta}}{\mathrm{sum}}& \left(\mathrm{Eq}.40\right)\\ d_0^p=d_0^p·\frac{\mathrm{delta}}{\mathrm{sum}}& \left(\mathrm{Eq}.41\right)\\ d_0^b=d_0^b·\frac{\mathrm{delta}}{\mathrm{sum}}& \left(\mathrm{Eq}.42\right)\end{array}$

[0183] Equations 40, 41, and 42 correspond to assignment statements for the values of the virtual buffer occupancy variables for the initial conditions d₀ⁱ, d₀^p, and d₀^b. The values to the right of the “=” correspond to the values before adjustment, and the values to the left of the “=” correspond to the values after adjustment. It will be observed that when the value for delta and the value for sum are approximately the same, that relatively little adjustment to the values occurs. When the value for sum is relatively high compared to the value for delta, the values of the virtual buffer occupancy variables for the initial conditions d₀ⁱ, d₀^p, and d₀^b are reduced proportionally. It should also be noted that relatively small values can also be added to the value of sum used in Equations 40-42 to prevent division by zero problems. After adjustment, the process ends by, for example, proceeding to the state 614 of the process described earlier in connection with FIG. 6.

[0184] FIG. 9B is a flowchart that generally illustrates a process for resetting virtual buffer occupancy levels upon the detection of an irregularity in a final buffer occupancy level. The process for resetting can be incorporated into encoding processes, such as in the state 627 of the rate control and quantization control process described earlier in connection with FIG. 6.

[0185] The process begins at a decision block 952. As explained earlier in connection with the state 627 of the rate control and quantization control process described in connection with FIG. 6, the final occupancy (fullness) of the applicable virtual buffer, i.e., the value of d_jⁱ, d_j^p, or d_j^b, where j=MB_cnt, can be used as the initial condition for the encoding of the next picture of the same type, i.e., as the value for d₀ⁱ, d₀^p, or d₀^b for the picture of the same type (I, P, or B). When encoding via the process described in TM5, the final occupancy of the applicable virtual buffer, i.e., the value of d_jⁱ, d_j^p, or d_j^b, is always used as the initial condition for the encoding of the next picture of the same type. However, the final occupancy of the applicable virtual buffer is not always an appropriate value to use.

[0186] In the decision block 952, the process determines whether the final occupancy of the applicable virtual buffer, i.e., the value of d_jⁱ, d_j^p, or d_j^b, is appropriate to use. In one embodiment, the appropriateness of a value is determined by whether the value is physically possible. A virtual buffer models a physical buffer. A physical buffer can be empty, can be partially occupied with data, or can be fully occupied with data. However, a physical buffer cannot hold a negative amount of data. To distinguish between physically attainable values and non-physically attainable values, one embodiment of the process compares the value for the final occupancy of the applicable virtual buffer to a predetermined threshold tr.

[0187] In one embodiment, the value of tr is zero to distinguish between a physically attainable buffer occupancy and a buffer occupancy that is not physically attainable. In one embodiment, a value that is relatively close to zero is used. Although the value of tr can correspond to a range of values, including values near to zero such as one, two, three, etc., the value of tr should not permit a negative value for the final occupancy to be deemed appropriate. It will be understood that when the value used for tr is zero, the process can distinguish between physically attainable values and non-physically attainable values by inspecting the sign, i.e., positive or negative, associated with the value of the final occupancy of the applicable virtual buffer. It will also be understood that when integer comparisons are made, a comparison using an inequality such as greater than negative one, i.e., >−1, can also be used, such that a value for tr can correspond to −1. The process proceeds from the decision block 952 to a state 954 when the final occupancy value is not appropriate to use as an initial condition for the next picture of the same type. Otherwise, the process proceeds from the decision block 952 to a state 956.

[0188] In the state 954, the process resets the final buffer occupancy value for the picture type that had just been encoded d_jⁱ, d_j^p, or d_j^b, where j=MB_cnt, to an appropriate value, such as a physically attainable value. Appropriate values can include any value from zero to the capacity of the applicable virtual buffer. In one embodiment, the final buffer occupancy value is reset to a relatively low value that is near zero, such as zero itself. The process can advance from the state 954 to an optional state 958, or the process can advance from the state 954 to the state 956.

[0189] In the optional state 958, the process normalizes the virtual buffer occupancy values d_jⁱ, d_j^p, and d_j^b. In the prior state 954, the process had corrected for a non-physically attainable value in the virtual occupancy value d_jⁱ, d_j^p, or d_j^b, that applies to the type of picture that was encoded. For example, the process can take the prior negative value of the applicable virtual occupancy value d_jⁱ, d_j^p, or d_j^b, and allocate the negative value to the remaining virtual occupancy values such that the sum of the virtual occupancy values d_jⁱ, d_j^p, and d_j^b, sums to zero. For example, in one embodiment, the process adds half of the negative value to each of the two other virtual occupancy values. The process advances from the optional state 958 to the state 956.

[0190] In the state 956, the process stores the final virtual buffer occupancy value as reset by the state 954 or unmodified via the decision block 952 and ends. The process can end by, for example, proceeding to the state 619 of the rate control and quantization control process described earlier in connection with FIG. 6.

Scene Change Within a Group of Pictures

[0191] FIG. 10A illustrates examples of groups of pictures. Scene changes between pictures of a sequence can exist within a group of pictures. Scene changes are relatively commonly encountered in a sequence of pictures. The scene changes can result from a change in camera shots, a switching between programs, a switch to a commercial, an edit, and the like. With a scene change, the macroblocks of a present picture bear little or no relation to the macroblocks of a previous picture, so that the macroblocks of the present picture will typically be intra coded, rather than predictively coded. Since an I-picture includes only intra-coded macroblocks, scene changes are readily accommodated with I-pictures.

[0192] Although pictures corresponding to scene changes are preferably coded with I-pictures, the structure of a group of pictures, i.e., the sequence of picture types, can be predetermined in some systems or outside of the control of the encoder. For example, one direct broadcast satellite (DBS) system has a predetermined pattern of I-pictures, P-pictures, and B-pictures that is followed by the encoder. As a result, scene changes can occur in B-pictures or in P-pictures. A conventional encoder can accommodate scene changes in B-pictures by referencing the predictive macroblocks of the B-picture to an I-picture or to a P-picture that is later in time.

[0193] A scene change in a P-picture can be problematic. A P-picture can include intra-coded macroblocks and can include predictively-coded macroblocks. However, a P-picture cannot reference a picture that is later in time, so that the scene change will typically be encoded using only intra-coded macroblocks. In substance, a scene change P-picture in a conventional encoder is an I-picture, but with the bit allocation and the header information of a P-picture. In a conventional encoder, a P-picture is allocated fewer bits than an I-picture so that the picture quality of a scene change P-picture is noticeably worse than for an I-picture. Other pictures, such as B-pictures and other P-pictures, can be predictively coded from the P-picture with the scene change, thereby disadvantageously propagating the relatively low picture quality of the scene change P-picture.

[0194] As described earlier in connection with FIGS. 1 and 5, the pictures of a sequence are arranged into groups of pictures. A group starts with an I-picture and ends with the picture immediately prior to a subsequent I-picture. The pictures within a group of pictures can be arranged in a different order for presentation and for encoding. For example, a first group of pictures 1002 in a presentation order is illustrated in FIG. 10A. An I-picture 1004 for a next group of pictures is also shown in FIG. 10A.

[0195] The pictures of a sequence can be rearranged from the presentation order when encoding and decoding. For example, the first group of pictures 1002 can be rearranged to a second group of pictures 1010, where the group is a first group of a sequence, and can be rearranged to a third group of pictures 1020, where the group is an ongoing part of the sequence. The second group of pictures 1010 and the third group of pictures 1020 are illustrated in encoding order. The end of the second group of pictures 1010 occurs when an I-picture 1012 from another group is encountered. Due to the reordering, two B-pictures 1014, 1016 that were originally in the first group of pictures 1002 in the presentation order are now no longer in the group of pictures as rearranged for encoding. With respect to the process described in connection with FIG. 10B, a group of pictures relates to a group in an encoding order.

[0196] The third group of pictures 1020 will be used to describe the process illustrated in FIG. 10B. The third group of pictures 1020 includes two pictures 1022, 1024 that will be presented before the I-picture 1026 of the third group of pictures 1020. In the illustrated example, a scene change occurs in the third group of pictures 1020 at a P-picture 1030 within the third group of pictures 1020. The process described in FIG. 10B advantageously recognizes the scene change and reallocates the remaining bits for the remaining pictures 1032 in the third group of pictures 1020 to improve picture quality.

[0197] FIG. 10B is a flowchart that generally illustrates a process for resetting encoding parameters upon the detection of a scene change within a group of pictures (GOP). In the illustrated embodiment of the process, the encoding order is used to describe the grouping of groups of pictures.

[0198] The process illustrated in FIG. 10B identifies scene-change P-pictures and advantageously reallocates bits within the remaining pictures of the group of pictures without changing the underlying structure of the group of pictures. The process advantageously allocates relatively more bits to the scene change P-picture, thereby improving picture quality. The illustrated process can be incorporated into the rate control and quantization control process described earlier in connection with FIG. 6. For example, the process of FIG. 10B can be incorporated before the state 610 of FIG. 6. The skilled practitioner will appreciate that the illustrated process can be modified in a variety of ways without departing from the spirit and scope of the invention. For example, in another embodiment, various portions of the illustrated process can be combined, can be rearranged in an alternate sequence, can be removed, and the like.

[0199] The process begins at a decision block 1052. In the decision block 1052, the process determines whether there has been a scene change or a relatively sudden increase in an amount of motion in a picture. The scene change can be determined by a variety of techniques. In one embodiment, the process makes use of computations of picture comparisons that are already available. For example, one embodiment of the process uses a sum of absolute differences (SAD) measurement. The SAD measurement can be compared to a predetermined value, to a moving average, or to both to determine a scene change. For example, a SAD measurement that exceeds a predetermined level, or a SAD measurement that exceeds double the moving average of the SAD can be used to detect a scene change. Advantageously, the SAD measurement can detect a scene change or a sudden increase in an amount of motion in a picture. It will be understood that there may be another portion of the encoding process that also monitors for a scene change, and in one embodiment, the results of another scene change detection is reused in the decision block 1052. The process proceeds from the decision block 1052 to a decision block 1054 when a scene change is detected. Otherwise, the process proceeds to end, such as, for example, entering the state 610 of the rate control and quantization control process described earlier in connection with FIG. 6.

[0200] In the decision block 1054, the process determines whether the type of the picture to be encoded corresponds to the P-type. In another embodiment, the order of the decision block 1052 and the decision block 1054 are interchanged from that shown in FIG. 10B. The process proceeds from the decision block 1054 to a state 1056 when the picture is to be encoded as a P-picture. Otherwise, the process proceeds to end by, for example, entering the state 610 of the rate control and quantization control process described earlier in connection with FIG. 6.

[0201] In the state 1056, the process reallocates bits among the remaining pictures of the group of pictures. Using the third group of pictures 1020 of FIG. 10A as an example, when a scene change is detected at the P-picture 1030, the remaining bits R are advantageously reallocated among the remaining pictures 1032. In one embodiment, the process encodes the remaining pictures 1032 as though the P-picture 1030 is an I-picture, but without altering the structure of the group of pictures by not changing the type of picture of the P-picture 1030.

[0202] The process for encoding the P-picture 1030 as though it is an I-picture can be performed in a number of ways. For example, one embodiment of the process effectively decrements the number of P-pictures N_p to be encoded before the P-picture with the scene change is encoded, and uses the decremented value of N_p in Equation 6 to generate a targeted bit allocation. Equation 6, which is used in a conventional system only to calculate a targeted bit allocation T_i for a I-picture, can be used by the process of FIG. 10B to calculate a targeted bit allocation for the P-picture with the scene change. Equation 43 illustrates an expression of such a targeted bit allocation, expresses as T_p′. $\begin{array}{cc}{T}_{p^{\prime }}=\max \left\{\left(\frac{R}{\left(1+\frac{\left(N_p-1\right)X_p}{X_i{K}_p}+\frac{N_b{X}_b}{X_i{K}_b}\right)}\right),\left(\frac{\mathrm{bit\_rate}}{8·\mathrm{picture\_rate}}\right)\right\}& \left(\mathrm{Eq}.43\right)\end{array}$

[0203] This advantageously allocates to the P-picture a relatively large number of bits, such that the P-picture with the scene change can encode the scene change with relatively high quality. Equations 7 and 8 can then be used for the subsequent encoding of P-pictures and B-pictures that remain to be encoded in the group of pictures. Optionally, the process can further reset the values for the complexity estimators X_i, X_p, and X_b in response to the scene change by, for example, applying Equations 1-3 described earlier in connection with the state 608 of the rate control and quantization control process of FIG. 6. The process then ends by, for example, proceeding to the state 610 of the rate control and quantization control process. It will be understood that the process described in connection with FIGS. 10A and 10B can be repeated when there is more than one scene change in a group of pictures.

Selective Skipping of Macroblocks in B-Pictures

[0204] FIG. 11 is a flowchart that generally illustrates a process for the selective skipping of data in a video encoder. This selective skipping of data advantageously permits the video encoder to maintain relatively good bit rate control even in relatively extreme conditions. The selective skipping of data permits the video encoder to produce encoded data streams that advantageously reduce or eliminate relatively low occupancy levels in a decoder buffer, such as decoder buffer underrun. Decoder buffer underrun can occur when the playback bit rate exceeds the relatively constant bit rate of the data channel for a sustained period of time such that the decoder buffer runs out of data. Decoder buffer underrun is quite undesirable and results in a discontinuity such as a pause in the presentation.

[0205] Even without an occurrence of decoder buffer underrun, data streams that result in relatively low decoder buffer occupancy levels can be undesirable. As explained earlier in connection with FIG. 4, a buffer model, such as the VBV buffer model, is typically used in an encoding process to model the occupancy levels of a decoder buffer. When a conventional encoder determines that the occupancy level of the buffer model is dangerously low, the conventional encoder can severely compromise picture quality in order to conserve encoding bits and maintain bit rate control. The effects of relatively low VBV buffer model occupancy levels is noticeable in the severely degraded quality of macroblocks.

[0206] The process generally illustrated by the flowchart of FIG. 11 advantageously skips the encoding of selected macroblocks when relatively low buffer model occupancy levels are detected, thereby maintaining relatively good bit rate control by decreasing the number of bits used to encode the pictures in a manner that does not impact picture quality as severely as conventional techniques. In one example, the process illustrated in FIG. 11 can be incorporated in the state 623 of the rate control and quantization control process described earlier in connection with FIG. 6. The skilled practitioner will appreciate that the illustrated process can be modified in a variety of ways without departing from the spirit and scope of the invention. For example, in another embodiment, various portions of the illustrated process can be combined, can be rearranged in an alternate sequence, can be removed, and the like.

[0207] The process starts at a decision block 1102, where the process determines whether the picture designated to be encoded corresponds to a B-picture. B-pictures can be encoded with macroblocks that are predictively coded based on macroblocks from other pictures (I-pictures or P-pictures) that are earlier in time or later in time in the presentation order. However, during the encoding process, the pictures (I-pictures or P-pictures) that are used to encode a B-picture are encoded prior to the encoding of the B-picture. The process proceeds from the decision block 1102 to a decision block 1104 when the picture to be encoded is a B-picture. Otherwise, the process proceeds to end, by, for example, returning to the state 623 of the process described earlier in connection with FIG. 6.

[0208] In the decision block 1104, the process determines whether the VBV buffer occupancy level is relatively low. During the encoding process, a relatively large number of bits may have already been consumed in the encoding of the pictures from which a B-picture is to be encoded. In some circumstances, this consumption of data can lead to a low VBV buffer occupancy level. For example, the process can monitor the occupancy level V_status of the VBV buffer model, which was described earlier in connection with FIG. 7, and compare the occupancy level V_status to a predetermined threshold, such as to V_critical. The comparison can be made in a variety of points in the encoding process. In one embodiment, the comparison is made after a picture has been encoded and after the VBV buffer model occupancy level has been determined, such as after the state 638 or after the state 610 of the rate control and quantization control process described earlier in connection with FIG. 6. In one embodiment, the comparison is advantageously made before any of the macroblocks in the picture have been encoded, thereby advantageously preserving the ability to skip all of the macroblocks in the picture when desired to conserve a relatively large amount of bits.

[0209] In one example, V_critical is set to about ¼ of the capacity of the VBV buffer model. It should be noted that the capacity of the VBV buffer model or similar buffer model can vary with the encoding standard. It will be understood that an appropriate value for V_critical can be selected from within a broad range. For example, other values such as 1/16, ⅛, 1/10, and 3/16 of the capacity of the VBV buffer model can also be used. Other values will be readily determined by one of ordinary skill in the art. In one embodiment, the process permits the setting of V_critical to be configured by a user. The process proceeds from the decision block 1104 to a state 1106 when the occupancy level V_status of the VBV buffer model falls below the predetermined threshold. Otherwise, the process proceeds from the decision block 1104 to a state 1108.

[0210] In the state 1106, the process skips macroblocks in the B-picture. In one embodiment, all the macroblocks are skipped. In another embodiment, selected macroblocks are skipped. The number of macroblocks skipped can be based on, for example, the occupancy level V_status of the VBV buffer. Data for an “encoded” B-picture is still formed, but with relatively little data for the skipped macroblocks. In the encoding process, a bit or flag in the data stream indicates a skipped macroblock. For example, in a technique known as “direct mode,” a flag indicates that the skipped macroblock is to be interpolated during decoding between the macroblocks of a prior and a later (in presentation time) I- or P-picture. Another flag indicates that the skipped macroblock is to be copied from a macroblock in a prior in presentation time I- or P-picture. Yet another flag indicates that the skipped macroblock is to be copied from a macroblock in a later in presentation time I- or P-picture. The skipping of macroblocks can advantageously encode a B-picture in relatively few bits. In one example, a B-picture for MPEG-2 with all the macroblocks skipped can advantageously be encoded using only about 300 bits. After the skipping of macroblocks for the B-picture is complete, the process ends by, for example, returning to the state 623 of the process described earlier in connection with FIG. 6.

[0211] In the state 1108, the process has determined that the occupancy level V_status of the VBV buffer is not relatively low, and the process encodes the macroblocks in the B-picture. After the encoding of the macroblocks for the B-picture is complete, the process ends by, for example, returning to the state 623 of FIG. 6. It will be understood that the decisions embodied in the decision block 1102 and/or the decision block 1104 can be performed at a different point in the process of FIG. 6 than the state 1106 or the state 1108.

Adaptive I, P-Frame and B-Frame DCT Filtering

[0212] An MPEG or other video encoder can encounter video sequences that are relatively difficult to compress at relatively low bit rates. This situation can be avoided by using variable bit-rate compression, where the encoder raises the instantaneous bit rate for these difficult to encode scenes. However, constant bit-rate compression techniques, which are encountered in Video Compact Disk (VCD), in digital TV broadcasting (including satellite, cable, and over the air), and the like, an accommodating increase in the bit rate is usually not available as VBV buffer model compliance constrains available the number of bits available to encode pictures. Instead, quality suffers. For example, conventional encoders increase the value of the quantization parameter QP, which then results in visible compression artifacts. Examples of compression artifacts that can be visually observed include blockiness and jerkiness in these difficult to encode video scenes.

[0213] One process described earlier in connection with FIG. 11 preserves compliance with VBV buffer requirements and reasonable picture quality by selectively skipping the encoding of macroblocks in B-pictures in response to VBV buffer model occupancy levels. The process described in FIGS. 12 and 13 can be used in combination with or independently of the process described earlier in connection with FIG. 11.

[0214] In the process of FIGS. 12 and 13, the process advantageously selectively and/or adaptively filters discrete cosine transform (DCT) coefficients to reduce the number of bits needed to encode pictures. The process can adaptively filter the DCT coefficients in response to VBV buffer model occupancy levels and to observed quantization parameter levels. The selective filtering advantageously improves the tradeoff between the number of bits used to encode a picture and the quality of the picture in situations where encoding bits are limited and VBV buffer model compliance or similar buffer model compliance is desired. Such filtering of DCT coefficients differs from the filtering provided in a conventional quantization table. The contents of a quantization table are typically fixed during encoding and are communicated in the bitstream such that the encoder and the decoder can use the same quantization table in a closed-loop manner. The filtering of the DCT coefficients described herein relates to a selective “open-loop” filtering of DCT coefficients only in the encoder. Advantageously, no compensating adjustments are needed in the decoder.

[0215] Selective DCT filtering will be generally described in the context of 8×8 image blocks. These 8×8 image blocks are used in MPEG1, in MPEG2, and in MPEG4. The skilled practitioner will appreciate that the DCT filtering techniques will also be applicable to other video compression standards, such as to H.264, where the block size is 4×4.

[0216] FIG. 12 is a flowchart that generally illustrates a process for selective DCT filtering. During the encoding process, the encoder performs a motion search 1210, performs motion compensation 1220, and proceeds to compute DCT coefficients 1230. In one embodiment, the process then selectively filters 1240 the DCT coefficients and/or selectively limits QP values 1250 to relatively low values. It will be understood that such filtering 1240 and limiting or bounding of QP values 1250 can be selectively applied in response to VBV buffer model occupancy levels and/or to the detection of relatively high QP values. While relatively high QP values can reduce the number of bits used to encode data, high QP values dramatically reduce picture quality by increasing picture blockiness. The process then proceeds to quantize the DCT coefficients 1260. When filtered and limited, the process quantizes the filtered DCT coefficients with the adjusted QP values. It will be understood that when VBV buffer model occupancy levels are relatively high, that the process can select not to filter the DCT coefficients and not to alter QP values. As will be described later in connection with FIG. 13, one embodiment of the process also varies the strength of the filtering in response to VBV buffer model occupancy levels. Prior to describing the process of FIG. 13, examples of filtering techniques in the frequency domain will first be described.

[0217] In the frequency domain, ω_x and ω_y denote the frequency domain variables along the horizontal and vertical directions after a 2-D Discrete Cosine Transform (DCT) of an 8×8 block of image data or residual error following motion compensation. In this frequency domain, values of ω_x=0 and ω_y=0 correspond to the DC coefficient, i.e., the average value of the 8×8 block of image data or residual error, while values of ω_x=7 and ω_y=7 correspond to the highest frequency details of the 8×8 block of image data or residual error. It will be understood that for a 4×4 block of image data, values of ω_x=3, ω_y=3 correspond to the highest frequency details.

[0218] Human vision and perception is relatively more sensitive to low frequencies and is relatively less sensitive to relatively high frequencies. At the same time, relatively many video images have less energy in high frequencies compared to that in low frequencies. Nevertheless, when video information is compressed by quantization of DCT coefficients, and especially when the Quantizer scale assumes high values (implying large quantization error), the residual error of high-detail frames has significant energy in the high frequency region. This results in these pictures requiring a significant amount of bits in order to be compressed, even at the highest Quantizer scale value.

[0219] In the process that will be described later in connection with FIG. 13, such difficult to encode scenes are detected and a low pass filter is selectively applied to image blocks during encoding. When applied, filtering is provided after performing motion search, after motion compensation, and after DCT computation, but before quantization. In this manner, high frequency coefficients are sacrificed (their value is reduced or even set to zero) and thus spatial resolution is somewhat reduced, especially in high-detail portions of a video picture. However, the overall balance in image quality is significantly improved. In one embodiment, a non-separable pyramid-shaped low-pass filter is used. “Pyramid” refers to the sloping shape of the filter response as it varies from low frequency (from DC) to high frequency (max AC). Of course, other filter shapes that can be used can be readily determined by one of ordinary skill in the art. In one embodiment, the pyramid-shaped filter depends on a positive valued parameter, s≧1, and one form can be expressed in Equation 44. $\begin{array}{cc}{H}_s\left({\omega }_x,{\omega }_y\right)=\frac{\max \left\{0,s-{\omega }_x-{\omega }_y\right\}}{s},0\le {\omega }_x,{\omega }_y\le 7& \left(\mathrm{Eq}.44\right)\end{array}$

[0220] It will be understood that the range for the frequency variables will depend on the size of the image block defined in the applicable encoding standard. The inverse of s, i.e., (1/s), relates to the strength of the filter. In one embodiment, filter strength (1/s) can take on a value between 0 and 1. Note that when the strength of the filter (1/s) is zero, i.e., s→∞, the response H_s has the constant value=1, which means that there is no filtering (DCT coefficients are not changed). When the strength of the filter (1/s) is relatively high, s=1, and H_s has the constant value of 0 everywhere except when ω_x=ω_y=0(DC coefficient). In this limiting case, application of the filter effectively down-samples the video picture by a factor of 8 in both the horizontal and the vertical direction.

[0221] The shape of the filter can vary with the strength parameters. In the illustrated examples, the strengths (1/s) described correspond to 1/32 (weak), 1/16 (medium) and ⅛ (strong). Tables III, IV, and V illustrate one example of a shape for these varying filter strengths. Graphs corresponding to the filters of Tables III, IV, and V are shown in FIGS. 14, 15, and 16, respectively. It will be understood by the skilled practitioner that many variations are possible. Both the shape of the filter can be varied and the strength of the filter can be varied. It will also be understood that the number of filters of different strength can be varied. For example, in other embodiments, more than three different filters can be used or fewer than three filters.

[0222]

TABLE III
32	31	30	29	28	27	26	25
31	30	29	28	27	26	25	24
30	29	28	27	26	25	24	23
29	28	27	26	25	24	23	22
								/32 (weak, strength = 1/32)
28	27	26	25	24	23	22	21
27	26	25	24	23	22	21	20
26	25	24	23	22	21	20	19
25	24	23	22	21	20	19	18

[0223]

TABLE IV
16	15	14	13	12	11	10	9
15	14	13	12	11	10	9	8
14	13	12	11	10	9	8	7
13	12	11	10	9	8	7	6
								/16 (medium, strength = 1/16)
12	11	10	9	8	7	6	5
11	10	9	8	7	6	5	4
10	9	8	7	6	5	4	3
9	8	7	6	5	4	3	2

[0224]

TABLE V
8	7	6	5	4	3	2	1
7	6	5	4	3	2	1	0
6	5	4	3	2	1	0	0
5	4	3	2	1	0	0	0
								/8 (strong, strength = 1/8)
4	3	2	1	0	0	0	0
3	2	1	0	0	0	0	0
2	1	0	0	0	0	0	0
1	0	0	0	0	0	0	0

[0225] In Tables III, IV, and V, the DC DCT coefficient (ω_x=0 and ω_y=0) is in the upper left. The horizontal frequency variable ω_x increases to the right, and the vertical frequency variable ω_y increases in the downward direction. Other strengths and shapes will be readily determined by one of ordinary skill in the art.

[0226] It should be noted that in the illustrated example, the filter weight uses the values of 9/16 ( 18/32) (weak), ⅛ (2/16) (medium) and 0 (strong), respectively, for the three filter strengths at the highest frequency DCT coefficient location, AC(7,7), while maintaining the value of 1 at the DC location (0,0). It should be noted that the filter should be determined such that the DC coefficient is not changed. Filter values for other DCT coefficients are located on a plane that connects these points, such that the higher the filter strength, the steeper the plane inclination is. Of course, other variations will be readily determined by one of ordinary skill in the art and can include shapes with curves.

[0227] In one embodiment, the variable s is selected to be a power of 2 (such as ⅛, ⅙, 1/32, etc) such that division can be efficiently accomplished by relatively fast bit shifting operations rather than complex division operations. However, it will be understood that the strength of the filter can assume other positive values between 0 and 1. Further advantageously, relatively many general purpose processors support the execution of multiple variable multiplication operations and division operations using binary shifts, such as the processors that support a Multimedia Extension (MMX) instruction set or the Single Instruction Multiple Data (SIMD), described earlier in connection with FIG. 6. This power of 2 utilization greatly simplifies overall implementation on these and similar architecture, and increases processing speed such that video encoding can be performed in real time. For example, the filtering process can advantageously be implemented with relatively negligible computational overhead, and in one example, was approximately 1%.

[0228] As will be described in greater detail later in connection with FIG. 13, the DCT filtering can be applied to the entire video sequence, adaptively on a case by case basis for a picture, on a macroblock-by-macroblock basis, or a combination of the three. It will be understood that the low-pass filter can be configured to always apply, such as, for example, in a dedicated hardware encoder, but that the filter response can be selected such that the low-pass filter has no effect, thereby effectively disabling the filter. A number of factors can be used to whether to apply the DCT filtering and the strength of DCT filtering should it be applied. In one embodiment, state variables that are typically already computed are advantageously used to make the decisions such that additional processor overhead and/or additional hardware circuitry is relatively low. For example, in one embodiment, the DCT filtering is selected based on observations of the behavior of the following two parameters: VBV buffer model occupancy and average running QP of current frame. The VBV buffer model was described in greater detail earlier in connection with FIG. 7. Advantageously, the theoretical occupancy level of the VBV buffer model or similar buffer model is already maintained by a video encoder, such as an MPEG video encoder.

[0229] The average running quantization parameter QP is also a state variable that is typically maintained in a video encoder. While many variations exist to calculate a running average, for example, how many samples are used in the running average, whether the running average is unevenly weighted, etc., many video encoders track an average value for the QP to monitor video quality. A running average QP that is relatively high, such as a QP that is close to an upper bound for the QP, e.g., close to 31, indicates relatively low quality encoding with a relatively blocky picture. By contrast, a relatively low average running QP, e.g., an average close to a lower bound such as 1, indicates that the encoder is encoding with relatively high video quality and without relatively much blockiness. In one embodiment, the number of pictures (N) typically used to calculate a moving or running average QP is in the range 2-30.

[0230] In one embodiment, in response to the VBV buffer model status and the average running QP, a tradeoff is made between (a) using DCT filtering and using a lower QP value or (b) not performing DCT filtering and using a higher QP value. It will be understood that (b) can correspond to conventional encoding techniques.

[0231] In one embodiment, when the encoder applies DCT filtering, another decision is made to determine the strength or magnitude of filtering. As will be understood by the skilled practitioner, the number of filter strengths that can be selected by the encoder can vary in a broad range, such as 2, 3, 4 or more filter strengths. In the illustrated embodiment, with Tables III, IV, and V, the encoder selects among 3 filter strengths ( 1/32=“weak”, 1/16=“medium” and ⅛=“strong”) with the corresponding filter coefficient matrices. Many variations on the filter strengths are possible.

[0232] In another example of filter coefficients, the amount of filtering performed by the filter can be varied by adjusting the frequency response of the filter as illustrated in Tables VI, VII, and VIII. The amount of filtering can thus be varied even if the strength (1/s) of the filter stays the same. In the filter responses described in Tables VI, VII, and VIII, the strength (1/s) of the filters is 1/16 for each filter.

[0233]

TABLE VI
16	15	14	13	12	11	10	9
15	14	13	12	11	10	9	0
14	13	12	11	10	9	0	0
13	12	11	10	9	0	0	0
								/16 (weak)
12	11	10	9	0	0	0	0
11	10	9	0	0	0	0	0
10	9	0	0	0	0	0	0
9	0	0	0	0	0	0	0

[0234]

TABLE VII
16	14	12	10	8	6	4	2
14	12	10	8	6	4	2	0
12	10	8	6	4	2	0	0
10	8	6	4	2	0	0	0
								/16 (medium)
8	6	4	2	0	0	0	0
6	4	2	0	0	0	0	0
4	2	0	0	0	0	0	0
2	0	0	0	0	0	0	0

[0235]

TABLE VII
16	8	4	2	0	0	0	0
8	4	2	0	0	0	0	0
4	2	0	0	0	0	0	0
2	0	0	0	0	0	0	0
								/16 (strong)
0	0	0	0	0	0	0	0
0	0	0	0	0	0	0	0
0	0	0	0	0	0	0	0
0	0	0	0	0	0	0	0

[0236] The filtering operation involves a simple scalar multiplication between the DCT coefficient and the coefficients listed previously. Many processors support instructions that perform multiplication of multiple variables in parallel. An example of such an instruction for a processor that supports the SSE instruction set or the MMX instruction set is an instruction for packed multiply, “PMUL.” Advantageously, the division operation can correspond to an efficient right bit-shifting operation as long as the factor is a power of 2. A process for adaptively selects DCT filtering will now be described.

[0237] FIG. 13 is a flowchart that generally illustrates an example of a process for adaptively selecting DCT filtering in a video encoder. It will be appreciated by the skilled practitioner that the illustrated process can be modified in a variety of ways without departing from the spirit and scope of the invention. For example, in another embodiment, various portions of the illustrated process can be combined, can be rearranged in an alternate sequence, can be removed, and the like.

[0238] The process starts at a state 1302, where the process retrieves the occupancy level for the VBV buffer model, which is typically updated prior to encoding of the picture. Advantageously, the occupancy level of the buffer model is already calculated in an encoding process. It will also be understood that the process does not have to retrieve the occupancy level until the occupancy level is used, but the data retrieve is shown in the state 1302 for clarity. The process advances from the state 1302 to a state 1304.

[0239] In the state 1304, the process calculates or retrieves the running QP average, which is also known as a “moving average.” In one embodiment, the running average is reused from an existing running average. In another embodiment, the running average is separately calculated. In yet another embodiment, a new running average for QP is calculated from an existing running average and a recent running average to weight the running average as expressed in Equation 45.
QP_x—avg=½*(QP_run+QP_x—prev)  (Eq. 45)

[0240] In Equation 45, QP_run corresponds to an average running QP retrieved from a system state variable, and QP_x^—prev corresponds to the QP of the previous picture of the same type (I, P, or B). The result of Equation 45, QP_x—avg can then be used in a comparison with a selected threshold, QP threshold of filtering (denoted as QP_flt—thr). The process advances from the state 1304 to a decision block 1306.

[0241] In the decision block 1306, the process compares the running average QP to a threshold, such as to a predetermined threshold. In one embodiment, QP_flt—thr is selected such that QP_flt—thr corresponds to a video quality threshold for an average QP, such as QP_x—avg, that indicates that blockiness in the video is likely to be apparent. It will be understood by one of ordinary skill in the art that the threshold can vary in a broad range and can depend on environmental conditions such as the resolution of the video, the frame rate of the video, and the contraints on the bit rate. In one embodiment, the process uses a smaller threshold for lower resolution video than for higher resolution value. In one embodiment, the process uses a threshold of QP_flt—thr=6 for VCD-size video (352×240) or smaller resolutions. For DVD, i.e., NTSC (720×480), or similar size video, a value of QP_flt—thr=14 can advantageously be used. Smaller resolution video, e.g., VCD, can be more sensitive to video blockiness because lower resolution video has a larger actual pixel size than higher resolution video. Other values can also be used, and still other values can be readily determined by one of ordinary skill in the art for different coding standards (MPEG4/H.264) or for different video frame sizes and frame rates.

[0242] In the illustrated process, the process proceeds from the decision block 1306 to a state 1308 when the running average exceeds the threshold (QP_x—avg>QP_flt—thr). Otherwise, the process proceeds from the decision block 1306 to a state 1312.

[0243] In the state 1308, the process sets a flag to indicate to the a decision block 1322 to filter all the macroblocks of the picture in a state 1324. It will be understood that variations of the process are possible, and in another embodiment of the process, fewer than all the macroblocks are filtered in response to the detection of the running average exceeding the threshold. Returning now to the illustrated embodiment, where all macroblocks are selected to be filtered in response to the running average exceeding the threshold, the process advances from the state 1308 to a state 1310.

[0244] In the state 1310, the process selects a filter strength to apply to the DCT coefficients. As described earlier, the effect of the filter can vary depending on the frequency response of the filter and on the strength (1s) of the filter. In the illustrated process, the VBV buffer model occupancy level retrieved in the state 1302 is used to determine the filter strength. In one embodiment, where the initial and long-term target for VBV buffer model occupancy is about ⅞ of the capacity of the VBV buffer model (⅞ of vbv_buffer_size), the following criteria is used to select the strength of the filter. A goal of ⅞ of the VBV buffer model capacity is a representative goal a video encoder. It will be understood that the criteria can vary in a wide range and can vary depending on the number of filter strengths from which the encoder selects and based on different goals and/or threshold for VBV buffer model occupancy.

[0245] In one embodiment, the lower the VBV buffer model occupancy, the stronger the strength of the applied filter. In one embodiment, where VBV_cur denotes VBV buffer model occupancy of the current picture or frame the filter strength is selected as follows. DCT filtering strength is set to weak when (VBV_cur> 6/8*vbv_buffer_size). Below the threshold of 6/8*vbv_buffer_size, DCT filtering strength is set to medium if (VBV_cur>⅜*vbv_buffer_size). Below the threshold of ⅜*vbv_buffer_size, DCT filtering strength is set to strong. The process advances from the state 1310 to a state 1320.

[0246] Returning now to the decision block 1306, when the running average is less than or equal to the threshold, (QP_x—avg<=QP_flt—thr), the process proceeds to the state 1312, where selected macroblocks can be filtered. In one embodiment, the process determines that should filtering be applied to a selected macroblock, that the filtering will be at a relatively weak strength such as the weak filter strength described earlier in connection with Table III. The process advances from the state 1312 to the state 1320.

[0247] In the state 1320, the process calculates the quantization parameter for a macroblock. The state 1320 can correspond to the start of a loop for processing the macroblocks of a picture. Various techniques can be used to calculated the quantization parameter (QP). In one embodiment, the QP is calculated using standard techniques described in TM5 from MPEG. In another embodiment, other techniques are used, such as the QP computation techniques described earlier in connection with FIG. 6. In a TM5-type of rate control, the quantization parameter (QP) of a particular macroblock (MB) i is determined by the virtual buffer status and spatial activity of the macroblock. If the bit allocation of the previous MB is surpassed, i.e., too many bits were consumed, the TM5-rate control can compensate for this by assigning a higher QP value to a subsequent macroblock (MB).

[0248] It should be noted that the TM5-type of rate control (among others) uses a method it calls “adaptive quantization” to assign a relatively smaller QP to a macroblock in a flat region (low texture) and a relatively higher QP for a macroblock in a busy or high-texture region. As a result, relatively many macroblocks are disadvantageously and consistently assigned relatively high QP values that lead to visually unpleasant artifacts such as blockiness. As will be described later, one embodiment of the process advantageously reduces the relatively high QP values to decrease the blockiness, while performing DCT filtering to reduce the number of bits consumed in the encoding process to comply with VBV buffer model constraints. The process advances from the state 1320 to the decision block 1322.

[0249] In the decision block 1322, the process determines whether to apply the filtering to all the macroblocks or to a selected macroblock. In the illustrated process, the process monitors the status of the flag set in the state 1308. When all the macroblocks are to be filtered, the process proceeds from the decision block 1322 to the state 1324. Otherwise, the process proceeds from the decision block 1322 to a decision block 1328.

[0250] In the state 1324, the process applies the selected DCT filter to the DCT coefficients. The number of bits is reduced by filtering by increasing the number of zeroes in the data, thereby decreasing the amount of data to be compressed. For example, the presence of additional zeroes can advantageously increase the number of zeroes encountered consecutively, which can be represented by a compact run length. The process advances from the state 1324 to a state 1326.

[0251] In the state 1326, the process places an upper bound on the QP values to alleviate blockiness. It has been observed that the blocky effect is usually more visible in an intra-macroblock than in an inter-macroblock (including forward, backward, and interpolated modes) for the same value of QP used. In one embodiment, a smaller QP bound is advantageously used for intra-macroblocks than for inter-macroblocks. An upper bound for an intra-macroblock and an upper bound for an inter-macroblock are denoted by QP_intra—thr and QP_inter—thr, respectively. In one embodiment, these bounds are adaptively selected according to the occupancy of the VBV buffer model according to Equation 46. $\begin{array}{cc}\left({\mathrm{QP}}_{\mathrm{intra\_thr}},{\mathrm{QP}}_{\mathrm{inter\_thr}}\right)=\{\begin{array}{cc}\left(12,14\right),& \mathrm{if}{\mathrm{vbv}}_{\mathrm{cur}}\ge 7/8*\mathrm{vbv\_buffer}\mathrm{\_size}\\ \left(12,16\right),& \mathrm{else}\mathrm{if}{\mathrm{vbv}}_{\mathrm{cur}}\ge 6/8*\mathrm{vbv\_buffer}\mathrm{\_size}\\ \left(14,18\right),& \mathrm{else}\mathrm{if}{\mathrm{vbv}}_{\mathrm{cur}}\ge 5/8*\mathrm{vbv\_buffer}\mathrm{\_size}\\ \left(14,20\right),& \mathrm{else}\mathrm{if}{\mathrm{vbv}}_{\mathrm{cur}}\ge 4/8*\mathrm{vbv\_buffer}\mathrm{\_size}\\ \left(18,24\right),& \mathrm{else}\mathrm{if}{\mathrm{vbv}}_{\mathrm{cur}}\ge 3/8*\mathrm{vbv\_buffer}\mathrm{\_size}\\ \left(24,31\right),& \mathrm{else}\mathrm{if}{\mathrm{vbv}}_{\mathrm{cur}}\ge 2/8*\mathrm{vbv\_buffer}\mathrm{\_size}\end{array}& \left(\mathrm{Eq}.46\right)\end{array}$

[0252] For example, if the current VBV buffer model occupancy level is between about 4/8 and ⅚ of the capacity of the VBV buffer model, then an upper bound for the QP is selected to be 14 and 20, respectively, for intra-macroblocks and for inter macroblocks. Other values for bounds for QP, that is, other values for QP_intra—thr and QP_inter—thr can readily be determined by one of ordinary skill in the art. The upper bound is applied to the QP calculated in the state 1320. When the calculated QP from the state 1320 is larger than the given bound, (QP_intra—thr, QP_inter—thr), the associated QP bound is used. The process advances from from the state 1326 to a state 1330, where the macroblock is encoded.

[0253] Returning now to the decision block 1328, at this point in the process, the process has determined that application of the DCT filtering of the state 1324 and the possible QP adjustment of the state 1326 is determined for a macroblock on a case by case basis. In one embodiment, the process proceeds from the decision block 1328 to the DCT filtering state 1324 and to the QP adjustment state when the running QP exceeds the thresholds described earlier in connection with Equation 46. When the running QP does not exceed the threshold, the process does not apply DCT filtering and does not adjust the QP and proceeds to the state 1330.

[0254] In the state 1330, the process encodes the macroblock. The encoding can be unfiltered via entry to the state 1330 from the decision block 1328, or filtered via entry of the state 1330 from the state 1326. The process advances from the state 1330 to a decision block 1332.

[0255] In the decision block 1332, the process determines whether it has finished encoding all the macroblocks in the picture. When there remain macroblocks to be encoded, the process returns to the state 1320 to continue processing. Otherwise, the process ends and can be restarted as desired for the encoding of a subsequent frame.

Macroblock Mode Decision

[0256] FIG. 17 is a flowchart that generally illustrates a process for selecting a mode for a macroblock. Conventional video encoding techniques inadequately determine the encoding mode for a macroblock. One embodiment according to the invention advantageously includes the coding cost for motion vectors into the mode decision, thereby enhancing the quality of the selection of the mode. One embodiment further uses an L1 norm in the decision, thereby advantageously reducing the computational complexity of taking the coding cost for motion vectors into the mode decision.

[0257] In a state 1710, the process performs a motion search. To determine the appropriate coding mode for a certain macroblock, a conventional encoder can calculate the sum of squares, which is also known as the L2-norm, as a measure for the residual error after motion search for each coding mode of the macroblock (Inter with motion compensation, Inter with zero motion vector, Field-inter with motion compensation etc.). The sum of squares is then compared to a sum of squares of the original pixel values. In a conventional system, the minimum sum of squares indicates the optimal coding mode.

[0258] One disadvantage to these sum of squares calculations is that the sum of squares calculations are relatively computationally expensive, and slow down the encoding process. For example, it has been observed that this type of mode-decision sum of squares calculation can consume up to about 10% of the overall encoding complexity of MPEG compression.

[0259] In a state 1720, the process computes or retrieves data for a sum of absolute differences (SAD) for the motion search. In one embodiment of the invention, a sum of absolute values, also known as L1-norm, of the residual error is used as the measure for mode decision. Further advantageously, the sum of absolute values of the residual error is actually equal to the sum of absolute differences (SAD) that is used during motion search to determine the best candidate macroblock for prediction. Accordingly, except for the L1-norm of the original pixels, the L1-norm of the residual error is readily available to the encoder without further computation. Even where calculated, the L1-norm calculation is significantly simpler and takes fewer CPU cycles to calculate compared to an 2-norm calculation, since no multiplication is required to generate the L1-norm. In addition, selected processors directly support useful instructions as described earlier in connection with FIG. 6.

[0260] Another disadvantage to the conventional approach to mode decision is that the conventional mode decision is determined purely by the distortion measure, i.e., the mode decision selects the mode with the lowest distortion. However, each of the modes are not equal in terms of consumed bits. For example, macroblock modes without motion vectors can be significantly smaller than other modes. As a result, the mode selected by conventional approaches can select a mode that does not provide relatively low distortion per bits consumed.

[0261] In many situations, a mode with a slightly larger distortion with a smaller Differential Motion Vector (DMV) or less motion vectors (the value of DMV will determine the used bits for coding this motion vector) can have better rate-distortion performance than a mode that exhibits the lowest distortion. As an example of the DMV, if the motion vector (MV) for a particular MB is (4,3) and the MV predictor for this MB is (3,6), then the DMV is (4−3,3−6)=(1, −3). In this case, the value of DMV to use for the modified SAD is |1|+|−3|=4.

[0262] In a state 1730, the process compensates for a coding cost for the motion vectors in the SAD computation by adjusting SAD_ORG to SAD_MOD. One embodiment uses a relatively simple decision rule to consider both rate of DMVs and their associated distortions. In one embodiment, since the decision rule is highly related to the picture and motion compensation (MC) type, the decision rule is selected according to whether a picture is B-picture or P-picture and whether the picture uses field prediction or frame prediction (for interlaced video and for progressive video). The actual value of the DMV, expressed in units of half-pixel, can be used for this purpose. The sum of the absolute values of the two components (horizontal and vertical) can be used in the value the following formulae for the decision rule.

[0263] In one embodiment, the cost associated with DMVs is translated into an equivalent SAD, such that the SAD of various modes can be more appropriately compared. In one embodiment, the DMV associated with each mode is first calculated, then DMV modified SADs of each mode are calculated, and then these modified SADs are used for the mode decision. For example, for a P-picture, the following rules identified in Rule 1 are used. Other rules will incorporating the principles and advantages described herein will be readily determined by one of ordinary skill in the art.

[0264] In the following formula, SAD_mod corresponds to the SAD value after DMV translation and used for mode decision, while SAD_org is the SAD value obtained from motion estimation. It will be understood that the rules expressed below can be readily modified to accommodate an L2-norm computation as well.

P-Picture—Progressive Sequence

[0265] For these macroblocks (MBs), a MB can be encoded in one of three types: Intra, Inter_No_MC, and Inter_MC. Rule 1:

• [0266] A. Inter_MC SAD is modified as: SAD_mod(Inter_MC)=SAD_org+min((DMV+1)*4, 32)
• [0267] B. Inter_No_MC is kept as original: SAD_mod(Inter_No_MC)=SAD_org.
• [0268] C. Intra MB is modified as SAD_mod(Intra)=SAD_org−128.
  P Frame—Interlaced Sequence

[0269] For these macroblocks, a MB can be encoded in one of four types: Intra, Inter_No_MC, Inter_MC_frame and Inter_MC_field. Inter_MC_field mode uses two motion vectors, such that there are two DMVs correspondingly. Rule 2:

• [0270] A. Inter_MC_frame SAD is modified as: SAD_mod(Inter_MC_frame)=SAD_org+min((DMV+1)*4, 32)
• [0271] B. Inter_MC_field SAD is modified as: SAD_mod(Inter_MC_field)=SAD_org+min(Σ_k=1²(DMV_k+1)*4, 64)
• [0272] C. Inter_No_MC is kept as original: SAD_mod(Inter_No_MC)=SAD_org.
• [0273] D. Intra MB is modified as SAD_mod(Intra)=SAD_org−128.
  B Frame—Progressive Sequence

[0274] For these macroblocks, a MB can be encoded in one of four types: Intra, forward, backward and interpolated modes. Rule 3:

• [0275] A. Forward SAD is modified as: SAD_mod(Forward)=SAD_org+min((DMV+1)*4, 32)
• [0276] B. Backward SAD is modified as: SAD_mod(Backward)=SAD_org+min((DMV+1)*4, 32)
• [0277] C. Interpolated SAD is modified as: SAD_mod(Interpolated)=SAD_org+min(Σ_K=1²(DMV_k+1)*4, 64) since it uses two motion vectors;
• [0278] D. Intra MB is modified as SAD_mod(Intra)=SAD_org−128.
  B Frame—Interlaced Sequence

[0279] For these macroblocks, a MB can be encoded in one of seven types: Intra, forward, forward_field, backward, backward_field, interpolated and interpolated_field modes.

• [0280] A. SAD of forward mode is modified as: SAD_mod(forward)=SAD_org+min((DMV+1)*4, 32)
• [0281] B. SAD of forward field mode is modified as: SAD_mod(forward_field)=SAD_org+min(Σ_K=1² (DMV_k+1)*4, 64) since it uses two motion vectors
• [0282] C. SAD of backward mode is modified as: SAD_mod(backward)=SAD_org+min((DMV+1)*4, 32)
• [0283] D. SAD of backward field mode is modified as: SAD_mod(backward_field)=SAD_org+min(Σ_K=1² (DMV_k+1)*4, 64) since it uses two motion vectors
• [0284] E. SAD of interpolated frame mode is modified as SAD_mod(interpolate_field)=SAD_org+min(Σ_K=1² (DMV_k+1)*4, 64) since it uses two motion vectors
• [0285] F. SAD of interpolated field mode is modified as: SAD_mod(Interpolate_field)=SAD_org+min(Σ_K=1⁴ (DMV_k+1)*4, 128) since it uses four motion vectors
• [0286] G. Intra MB is modified as SAD_mod(Intra)=SAD_org−128.

[0287] In a state 1740, the process selects the mode for the macroblock. After the SAD calculations are translated and compared in accordance with the appropriate rule, one embodiment of the process selects the mode with the lowest SAD after translation. Advantageously, experimental results indicate that using the L1-norm and modified formulas for coding mode decision works better than the traditional mode decision that uses the L2-norm by an average of 0.4 dB in PSNR (peak signal-to-noise ratio), which represents a relatively significant quality improvement. In addition, the improvement in quality can also be accompanied by faster processing where the L1-norm is used rather than the L2-norm.

[0288] FIG. 18 is another flowchart that generally illustrates a process for selecting a mode for a macroblock. In a state 1802, motion estimation for a frame or picture is performed. The motion estimation is provided to a state 1804 to calculate differential motion vectors and provided to 1810 to adjust SAD calculations. In the state 1804, the differential motion vectors are calculated. It will be understood that the differential motion vectors depend on prior motion vectors, data from which is provided from 1806. In a state 1808, translated SAD values are computed. In a state 1812, the translated SAD values are compared, and the macroblock mode is selected. In one embodiment, the mode selected corresponds to the mode with the lowest overall SAD. The selected mode is stored in 1814. Via state 1816, additional macroblocks can be processed.

[0289] Various embodiments of the invention have been described above. Although this invention has been described with reference to these specific embodiments, the descriptions are intended to be illustrative of the invention and are not intended to be limiting. Various modifications and applications may occur to those skilled in the art without departing from the true spirit and scope of the invention as defined in the appended claims.