Buffer management in variable bit-rate compression systems

[0001] This is a continuation of application Ser. No. 08/366,339, filed Dec. 28, 1994 now abandoned.

[0002] The present invention relates to a method and apparatus for the compression of digital video signals and in particular to the use of such method and apparatus in the encoding and decoding of signals.

[0003] Variable bit rate (VBR) video compression is known to give advantages over constant bit-rate (CBR) video compression. The main reason for this is that in a CBR system the bit rate has to be set so that the worst case quality is acceptable while in a VBR system the bit rate is set so that the average quality (which is kept constant) is acceptable. The difference between the average bit rate of a VBR compressed signal compared to the bit rate of a CBR compressed signal has been found to be to be close to 30% in favour of VBR compression.

[0004] Though VBR compression is better than CBR, it can only be used in a limited number of applications. In principle the medium should be able to convey variable bit-rate signals. In terms of the ISO/OSI model, almost every medium conveys a fixed bit rate at the physical layer. On a higher level, the medium can be converted into a (logically) variable bit rate medium.

[0005] One particular application of VBR compression is joint bit-rate control, in which a number of sources make use of a single channel. In the case of joint bit-rate control the bit-rate of the video signals is controlled such that the individual bit rates can be varying (through optimal allocation of bit rate) but so that the sum of all bit rates is constant. This type of system may occur in for instance cable television or in satellite television services.

[0006] The idea of joint bit-rate control for multi-program video signal encoding has been found to be advantageous, particularly for video signals coded according to the ISO MPEG standards. Basically a system for joint bit-rate control needs technical measures for two problems, namely the bit allocation and the buffer management. To deal with bit allocation, the bit need of the programs is measured and the bits are spread accordingly over the programs. Buffer management encompasses several sub-problems, some of which are described in “Constraints on variable bit-rate video for ATM-networks” by Amy Reibman and Barry Haskell; in IEEE Transactions on Circuits and Systems for Video Technology, Vol 2, No. 4 December 1992 pp. 361-372. The Reibman and Haskell paper examines the constraints resulting from encoder and decoder buffering in an asynchronous transfer mode (ATM) network, in particular the additional constraints needed to prevent overflowing or underflowing of the decoder buffer when a variable bit-rate data channel links encoder and decoder. They describe a method in which the number of encoded bits for each video frame and the number of bits transmitted across the variable bit-rate channel are selected jointly, a necessity imposed by the differing constraints imposed on the transmitted bit rate by the encoder and decoder buffers respectively.

[0007] In all these applications the transmission of video can be in the MPEG format. An MPEG decoder contains a physical buffer, and correct MPEG bit streams must fulfil the video buffering verifier (VBV) constraints, which means that the signals may not overflow or underflow a hypothetical decoder buffer. As will be shown hereinafter, an incorrect buffer management may limit the performance of a VBR compression system. The analyses of buffering systems rely strongly on the concept of a system delay whereby, in order to have a continuous display of video, the system delay must be constant. This requirement has an effect on the buffering strategy.

[0008] For CBR systems, given a fixed decoder buffer size, the delay must be relatively large for a low bit-rate and relatively low for a high bit-rate in order to make full use of the available buffering space in the decoder. If the bit rate is variable a compromise setting of the buffering system must be used. As a consequence of this compromise the effective buffer size in the decoder is too low for an acceptable performance at low bit rates. Essentially, two things are required, namely a constant end-to-end delay, and a low buffering delay for high bit rates and a high buffering delay for low bit rates. These requirements would appear to be incompatible.

[0009] It is an object of the present invention to provide improved stability of decoder buffer operation.

[0010] It is a further object of the present invention to provide greater efficiency in encoder buffer management.

[0011] In accordance with the present invention there is provided a method of compression for transmission of encoded digital video signals having a variable number of data bits per image frame, comprising the steps of:

• • [0012] a) detecting a first bit rate of the encoded digital video signal bit stream;
  • [0013] b) sequentially writing the signal bit stream into a buffer at the first bit rate;
  • [0014] c) deriving a second bit rate as a percentage of the first bit rate, changes are which percentage are inversely relation to changes in the first bit rate;
  • [0015] d) reading the bit stream out of the buffer at the second bit rate.

[0016] By the provision of the “tunable delay” resulting from the varied output and input bit rates in the encoder buffer, the contents of the nominal decoder buffer will remain substantially constant. The relationship between input and output rates means that at high input signal bit rates, where the remainder of the buffering system (provided by the encoder stage and the decoder) provides a relatively low delay, the tunable delay provides a relatively high delay. At low input signal bit rates, the tunable delay is reduced to counter the relatively higher delay introduced by the remainder of the buffering system.

[0017] The range of possible input bit rates may suitably be divided into discrete levels with a derived second bit rate being maintained constant until an input signal bit rate within a different one of the levels is detected at which point the second bit rate is recalculated. For a minimum input bit rate, the second bit rate is suitably set to equal that bit rate: in other words, at minimum input bit rate, the delay is set to zero.

[0018] Successive groupings of one or more image frames may be specified with the second bit rate being derived from the first frame of the grouping and maintained constant until the bit rate of the first frame of a succeeding grouping is detected. The groupings may be determined on the basis of successive frames within a range of bit rates as mentioned above or may be specified on other criteria. For signals encoded according to the MPEG standard, a first grouping may be specified for I-pictures, with one or more other groupings for other types of image.

[0019] Also in accordance with the present invention there is provided a video signal encoding apparatus operable to encode a digital video signal for transmission, the apparatus comprising an encoder stage operable to encode a received video signal according to a predetermined coding scheme and to output the signal as a variable bit-rate data stream and a buffer coupled to receive the said variable bit-rate data stream from the encoder and arranged to output a data signal for transmission; characterised by means operable to detect the bit-rate of the said variable bit-rate data stream, to derive a second bit rate as a percentage of the encoder stage output bit rate, which percentage changes in inverse relation to changes in the encoder stage output rate, and to control the buffer output data signal bit rate at the said second bit rate.

[0020] Further in accordance with the present invention there is provided an encoded video signal compressed (for transmission) by the method recited above. The signal, comprising a bit stream of encoded data for a succession of image frames, has its instantaneous bit rate inversely related to the bit density of an image frame n frame periods later where n is determined by the said bit density. As will be understood, the bit rate of the uncompressed signal is determined by the bit density, and the n frame periods (where n is not necessarily an integer) corresponds to the lag introduced by the tunable delay.

[0029] The following description is concerned with the management of MPEG-coded video signals by way of example, although it will be readily appreciated by the skilled practitioner that the invention is not restricted to such coding standards.

[0030] The MPEG standard describes a syntax and semantics of bit streams for compressed video and associated audio. Whilst the semantics in principle specify the functionality of the decoder, the standard provides no prescriptions for the decoder architecture. Each decoder will have an input buffer of some sort but how this is to be realized and what exact size this buffer must have is not specified.

[0031] Recognising the importance of random access for stored video and the significant bit-rate reduction that may be obtained through motion-compensated interpolation, the MPEG standard recognises three types of pictures (frames) namely intra-pictures, predicted pictures and interpolated pictures, generally referred to as I, P, and B-pictures respectively. I-pictures provide access points for random pictures and accordingly have only limited compression. P-pictures are coded with reference to a past (I or P) picture and often form a reference for future P-pictures. B-pictures have the highest degree of compression but require both a past and future reference for prediction. A typical MPEG sequence of I, P and B-pictures is shown in FIG. 1.

[0032] A schematic representation of an MPEG encoder is shown in FIG. 2. A received video signal is passed to encoder stage 10 together with a control signal which varies the quantisation coarseness of the encoding. In many cases (although not for joint bit rate encoding) the control signal is kept constant to maintain coarseness at the same average level. The control signal is derived on the basis of the incoming video signal by detector 12. The output of the encoder stage 10 is passed to an encoder buffer 14 at a first bit rate and also to a calculation unit 16. The calculation unit 16 derives a value for a second bit rate and passes this to a header insertion stage 18 which reads the signal bit stream out of the buffer at the second bit rate and places the bit stream on a communication channel for transmission to a receiver/decoder. The basis of the calculations performed by unit 16 to derive an appropriate second bit rate will now be described in terms of a theoretical encoder/decoder buffer arrangement.

[0033] The MPEG standard constrains bit streams such that they may not overflow or underflow a hypothetical buffer: this hypothetical buffer can be related to the physical buffers appearing in the signal encoder and decoder stages.

[0034] In studies of buffer management problems, analysis of the combination of the encoder buffer and the decoder buffer uses a model of the two buffers as shown in FIG. 3: this model will be used to explain the problem that occurs in a variable bit-rate situation.

[0035] The MPEG standard is partitioned in three parts: the video part, the audio part and the systems part. The video part contains a mechanism called the VBV and the systems part contains a part called the transport system target decoder (T-STD). MPEG decoders are realisable because the signal (bitstream) has to fulfil certain conditions, and bitstreams that fulfil the VBV condition (or similar for T-STD) are said to be compliant. Since bitstreams must fulfil the VBV condition it is possible to design decoders that can decode compliant bitstreams.

[0036] The VBV definition is based on a hypothetical decoder which reads the compressed pictures in zero time from the VBV buffer and a channel that continuously fills up the buffer. FIG. 4 shows characteristic VBV buffer contents as a function of time for a part of the MPEG frame sequence of FIG. 1. The VBV is specifically defined for constant bit-rate operation. In case of variable bit-rate operation, the MPEG standard specifies that the definition of the VBV buffer is overruled by the definition of the T-STD.

[0037] The bitstream is written into the buffer at the bit-rate specified in the sequence header. The bitstream contains information about the contents of the VBV buffer. This information is presented in a field in each picture header called the VBV delay, which specifies the time between the moment that the header enters the VBV buffer and the moment of decoding the picture.

[0038] In the systems part of the MPEG standard, mechanisms are defined that provide means to synchronise the decoder. These mechanisms are called the decoding time stamp (DTS) and the presentation time stamp (PTS) which are measured in time units. The mechanism of system clock reference (SCR) is used to recover the correct time. The complete decoding operation is described again in terms of a hypothetical decoder which is referred to as the standard target decoder (STD). The systems part describes two types of systems layer: the transport stream and the program stream. The STD for the transport stream is referred to as T-STD and for the program stream the term P-STD is used. In the following the difference between these streams is explained.

[0039] In principle the program stream is defined for applications which have a very low error probability such as optical recording. The basic characteristic of the program stream is that it can convey only one program and that the bit-rate can be varying: as such it may be of limited general use. Of greater utility is the transport stream, which is intended for multi-program environments. The bit-rate for each program is allowed to vary but the sum of all bit-rates must be constant. The transport stream uses a fixed size packets of 188 bytes. The T-STD can be used for HD-CD (VBR optical medium) by reading a variable bit-rate from the disk and adding an artificial empty program to the bitstream in passing the information from the reader on to the decoder such that the total bit-rate is constant. The definition of the functionality of both the P-STD and the T-STD is similar.

[0040] We have recognised that a problem in conventional buffer-management is that the available buffer space in the decoder depends on the bit rate. At low bit rate we have too small an effective decoder buffer capacity available, which will hamper the image quality.

[0041] Referring back now to the model of FIG. 3, the buffering model uses the following parameters

• • [0042] the (discrete) time n
  • [0043] the current bit rate R[n] of readout from the encoder buffer
  • [0044] the input-output delay d
  • [0045] physical encoder buffer size B_E
  • [0046] physical decoder buffer size B_D
  • [0047] encoder buffer pointer E[n]
  • [0048] decoder buffer pointer D[n]
  • [0049] range of decoder buffer pointer D_min[n]≦D[n]<D_max[n]
  • [0050] range of encoder buffer pointer E_min[n]≦E[n]<E_max[n]

[0051] It will be noted that the range of the buffer pointers is introduced as a separate variable which is time dependent, with the current buffer pointers E[n] and D[n] representing the buffer contents. A key variable in the buffer system is the input-output delay d which denotes the time between inputting a picture to the encoder and the time that this picture is decoded by the decoder. In the following derivations, we assume d to be an integer number of pictures, although non-integer values are possible as well. This variable is necessarily constant over the entire sequence since otherwise a non-continuous display would result, affecting motion portrayal.

[0052] We assume that the buffers are operated in the following way
E[n]=E[n−1]−R[n]+p[n]
D[n]=D[n−1]+R[n]−p[n−d]  (Equation 1)

[0053] In this model we assume an action (next n) every new picture. From this buffer behaviour model the following relation between the encoder buffer contents and the decoder buffer contents can be derived $\begin{array}{cc}D\left[n\right]+E\left[n-d\right]=\sum _{i=n-d+1}^{n}R\left[i\right]& \left(\mathrm{Equation}2\right)\end{array}$

[0054] From Equation 2 we can derive relation between the boundaries of the buffer contents which yields $\begin{array}{cc}{D}_{\max }\left[n\right]=\sum _{i=n-d+1}^{n}R\left[i\right]-E_{\min }\left[n-d\right]D_{\min }\left[n\right]=\sum _{i=n-d+1}^{n}R\left[i\right]-E_{\max }\left[n-d\right]& \left(\mathrm{Equation}3\right)\end{array}$

[0055] This equation shows that underflow in the encoder is linked to overflow in the decoder and vice versa; the encoder buffer and decoder buffer have dual behaviour. Clearly, there is a lag of d samples between encoder and decoder constraints.

[0056] A common choice (that is made in the Reibman and Haskell paper mentioned previously) is to select:
D_min[n]=E_min[n−d]=0  (Equation 4)

[0057] This choice means that underflow in either of the two buffers is in fact a physical underflow and, hence, a loss of data. Underflow in the encoder means the following violation on the decoder side of
D[n]>D_max[n]  (Equation 5)
with $\begin{array}{cc}{D}_{\max }\left[n\right]=+\sum _{i=n-d+1}^{n}R\left[i\right]& \left(\mathrm{Equation}6\right)\end{array}$

[0058] Hence the maximum available buffer space in the decoder depends on the (integrated) bit rate over the past d samples. This may lead to problems if, for instance, the bit rate is constant over more than d samples. In such a situation we have
D_max[n]=d·R  (Equation 7)
hence at a low bit-rate we have a small effective decoder buffer available. At a low bit-rate we still may have relatively large I-pictures, which can happen for scenes with high detail and absolutely no motion. Since the available buffer space is small, it is a conventional technique to reduce the quality of that I-picture such that the buffer constraint in Equation 5 is not violated.

[0059] Our solution to this problem is to have a small delay, as shown in FIG. 5 and also to use a relatively high bit-rate for the I-picture. The small delay averages out over a small number of pictures, and hence a small bit-rate peak will help. The special treatment of I-pictures means that over a series of pictures such as the sequence of FIG. 1, we use two bit-rates; a relatively high one for the period of time containing the I-pictures and a lower one for the other classes of pictures. This provides us with higher peak bit-rates. For the sequence of FIG. 1, we therefore have three groups of pictures (GOP's), one for each of the I-pictures and one for the GOP of P- and B-pictures between them, with the bit-rate remaining constant (either high or low) in each GOP. If desired, a third bit-rate (between the relatively high and relatively low levels) may be specified for P-pictures in which case the sequence of FIG. 1 would comprise five GOP with the bit-rate following the pattern of high-low-medium-low-high over the series.

[0060] The use of this specified bit-rate technique in the encoder may be detected from the decoder by testing Equation 5 on the bit stream. If the equation yields an “overflow” it shows the technique to have been used.

[0061] The choices for our larger encoder buffer and appropriate bit-rate over time will now be described. The conventional choice made in Equation 4 is the source for existing problems: we have appreciated that by making an alternative choice more flexibility in the decoder is provided. We choose
D_min=0
D_max=B_D  (Equation 8)
which may be considered as an advance selection of the largest possible decoder buffer. As will be shown, this choice leads to a larger encoder buffer. Note that an alternative choice for the buffer, such as $\begin{array}{cc}{D}_{\max }\left[n\right]=A+B·\left(\sum _{i=n-d+1}^{n}R\left[i\right]\right)& \left(\mathrm{Equation}9\right)\end{array}$
will yield similar conclusions for the encoder buffer, unless A=0, B=1 (which is the case in the original problem).

[0062] Substituting the choices from Equation 8 in Equation 3 yields $\begin{array}{cc}{E}_{\min }\left[n-d\right]=\sum _{i=n-d+1}^{n}R\left[i\right]-B_D\mathrm{and}& \left(\mathrm{Equation}10\right)\\ E_{\max }\left[n-d\right]=\sum _{i=n-d+1}^{n}R\left[i\right]& \left(\mathrm{Equation}11\right)\end{array}$

[0063] These equations show that the lower limit of the encoder buffer contents is non-zero which means that the encoder buffer must be larger than is conventional. In fact the physical size of the encoder buffer equals that of the decoder buffer plus some margin. The margin is calculated from the maximum and minimum bit rate on the channel as will be demonstrated below.

[0064] Considering Equation 10 for our analysis, analysis of the extremal values of E_min yields insight in important system characteristics. One end of the scale is reached when E_min=0. This situation will occur if we transmit for more than d symbols at the minimum bit rate. Substituting these assumptions in Equation 10 yields
B_D=d·R_min  (Equation 12)

[0065] Thus we use the minimum bit rate to select the required delay. It is appreciated that this delay will generally be larger than in the conventional situation, but the actual value remains acceptable, especially in the case of recording applications.

[0066] At the other end of the scale we find the required extra buffer size by considering that the encoder buffer size can be written as
B_E=B_D+margin  (Equation 13)
where the margin is equal to the maximum value of E_min. The maximum value is reached if the bit rate is at its maximum value for more than d samples (pictures). In this case we find $\begin{array}{cc}\mathrm{margin}=d·R_{\max }=B_D·\frac{R_{\max }}{R_{\min }}& \left(\mathrm{Equation}14\right)\end{array}$

[0067] Turning now to the calculation of the time dependent bit rate R[n], the decoder buffer content is required to be the same at the beginning and at the end of a GOP. FIG. 4 illustrates this condition using the decoder buffer fullness.

[0068] From Equation 1 it can be seen that this condition means $\begin{array}{cc}\sum _{i=n-d+1}^{n}p\left[i-d\right]=\sum _{i=n-d+1}^{n}R\left[i\right]& \left(\mathrm{Equation}15\right)\end{array}$
and keeping the bit rate R[i]=R constant over a GOP of length N pictures we have $\begin{array}{cc}\sum _{i=n-d+1}^{n}p\left[i-d\right]=N·R& \left(\mathrm{Equation}16\right)\end{array}$

[0069] From Equation 15 it will be appreciated that the channel bit rate has to change d samples after the source bit rate has changed and also that the bit rate will generally have to be changed within a picture, rather than at the header of a picture, although this may occasionally occur. The bit rate changes are shown in terms of the MPEG syntax in FIG. 6.

[0070] The MPEG transport stream is meant for multi-program environments and the program stream is meant for environments with one program, like optical recording. FIG. 6 shows how a video bit stream can be wrapped in a program stream. In the program stream we have the MPEG concepts PACK-headers and PES-headers (PES-packetized elementary stream). The PACK-header contains information on the elementary stream rate (=ES-rate). Since PACK-headers are not necessarily constrained to precede picture-headers or PES-headers we are allowed to change the bit rate at any moment.

[0071] The consequences of Equation 15 and Equation 16 are shown in FIG. 7 which shows the skew between moment of inputting the groups of pictures to the buffer and the moment of changing the bit rate: the continuous lines represent the time interval in which the bits resulting from compressing a GOP are written into the buffer, and the dashed lines indicate the time interval for which the output bit rate calculated for that GOP is applied.

[0072] In the foregoing, we have shown an enhanced buffer management strategy that does not put constraints on the quality of compression, for only a small extra delay.

[0073] From reading the present disclosure, other modifications will be apparent to persons skilled in the art. Such modifications may involve other features which are already known in the design, manufacture and use of digital video signal encoding and decoding systems and devices and component parts thereof and which may be used instead of or in addition to features already described herein.