US 6,460,153 B1Grant

Apparatus and method for unequal error protection in multiple-description coding using overcomplete expansions

Issue Date:2002-10-01

•23 Claims

•17 Drawing Sheets

Abstract

A projection onto convex sets (POCS)-based method for consistent reconstruction of a signal from a subset of quantized coefficients received from an N×K overcomplete transform. By choosing a frame operator F to be the concatenization of two or more K×K invertible transforms, the POCS projections are calculated in R^Kspace using only the K×K transforms and their inverses, rather than the larger R^Nspace using pseudo inverse transforms. Practical reconstructions are enabled based on, for example, wavelet, subband, or lapped transforms of an entire image. In one embodiment, unequal error protection for multiple description source coding is provided. In particular, given a bit-plane representation of the coefficients in an overcomplete representation of the source, one embodiment of the present invention provides coding the most significant bits with the highest redundancy and the least significant bits with the lowest redundancy. In one embodiment, this is accomplished by varying the quantization stepsize for the different coefficients. Then, the available received quantized coefficients are decoded using a method based on alternating projections onto convex sets.

Metadata

Assignee

Microsoft Corp.

Inventors

Philip A. Chou
Sanjeev Mehrotra
Albert S. Wang

Application Information

Application Number:US 09/276,955

Filing Date:1999-03-26

Priority Date:1999-03-26

Art Unit:7

Classifications

IPC:

H03M 1300H04N 730H04N 764

Field of Search:

7147467147763484033484053488451

Patent Drawings (17 sheets)

Description

[0002] This application is related to co-pending application having Ser. No. 09/276,842 entitled RECONSTRUCTION OF MISSING COEFFICIENTS OF OVERCOMPLETE LINEAR TRANSFORMS USING PROJECTIONS ONTO CONVEX SETS and assigned to the same assignee as the present application and filed on the same day herewith, and which is incorporated by reference.

FIELD OF THE INVENTION

[0003] This invention relates generally to the field of computer data recovery transformations, and in particular to recovery of signal data such as image data, wherein portions of the data have been lost.

COPYRIGHT NOTICE/PERMISSION

[0004] A portion of the disclosure of this patent document contains material that is subject to copyright protection. The copyright owner has no objection to the facsimile reproduction by anyone of the patent document or the patent disclosure as it appears in the Patent and Trademark Office patent file or records, but otherwise reserves all copyright rights whatsoever. The following notice applies to the software and data as described below and in the drawing hereto: Copyright @ 1998-1999, Microsoft Corporation. All Rights Reserved.

BACKGROUND

[0005] Channels used to transmit data are sometimes “noisy” and can corrupt or lose portions of the data transmitted. A channel is said to “corrupt” portions of the data if it changes some data values during transmission without notice to the receiver. A channel is said to “lose” or “erase” portions of the data if it does not transmit some data values to the receiver, but notifies the receiver as to which data items were not transmitted. Accordingly, channels may be classified as “corruption” or “erasure” channels. An example of a corruption channel is a wireless communications system wherein data are occasionally corrupted due to electromagnetic noise in the atmosphere. Another example of a corruption channel is a storage system such as a magnetic disk, a Compact Disk Read-Only Memory (“CDROM”), or a Digital Versatile Disk (“DVD”), wherein data are occasionally corrupted due to scratches and dust on the recording media or due to writing or reading errors in the channels which move data onto or off from the media. An example of an erasure channel is a packet communications network wherein certain packets are lost in transmission, and wherein the receiver can detect which data are lost using packet sequence numbers and/or other means. Another example of an erasure channel is a storage system wherein the data are striped across a number of disks, any of which can fail at random, and whose failure can be detected, such that which data items are missing is known. Any corruption channel can appear, at a high level, to be an erasure channel, by appropriate low-level error detection processing. For example, a parity check or checksum word can be transmitted along with a block of data; if the received checksum does not agree with the received data, then an error is detected in the block of data, and the block of data can be considered erased by the channel. It is desirable to have ways to recover the data that is corrupted or lost. For example, image data that have either pixels or blocks that have been lost would be blank (perhaps white or black on a display monitor) or would appear to have “noise” or snow in the received images. A “pixel” is an individual picture element, and is generally represented by a single value indicating, for example, the intensity of image at a point, or the intensity of each of three colors. A “block” is a group of adjacent pixels, generally a rectangle. For example, a block could be a 16-by-16-pixel portion of a larger image.

[0006] A “channel” can be a transmission channel, such as digital television broadcast or cable television transmission, or a telephony packet network that moves data over a distance or space. A “channel” can also be a storage channel that stores the data on a storage medium and then reads the data at a later time.

[0007] Error-correction codes have been devised for both corruption and erasure channels, and are well known. (See S. Lin and D. J. Costello, Error Control Coding: Fundamentals and Applications, Prentice-Hall, 1983, herein Lin and Costello.) By far, the most common error-correction codes are based on linear transformations of the data to add redundancy. If x is a data vector with K elements, and F is an N×K matrix, with N′≧K, then y=Fx is a vector with N elements. N−K of these elements can be considered redundant information. By transmitting the code vector y instead of x, the receiver can recover x even if some elements of y are corrupted or lost. In particular, if y is transmitted through an erasure channel, and no more than N−K elements of y are erased, then x can be recovered perfectly. For example, if x is a vector containing 3 binary elements, and

F = [\begin{matrix} 1 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & 0 & 1 \\ 1 & 1 & 1 \end{matrix}],

[0008] then y is a vector containing 4 binary elements, the first three of which are equal to x (this is called the “systematic” part of y) and the last of which is equal to the sum of the bits in x, modulo 2. That is, the last bit of y is equal to the exclusive-OR, or parity, of x. If any three bits of y are recovered, then x can be recovered perfectly. On the other hand, if two or more bits of y are erased, then x cannot be recovered perfectly; the bits of x that are lost are unrecoverable in this case.

[0009] Error-correction codes are intended to recover symbolic data. In the example above, the elements of x are binary symbols. In general, they can be members of a finite field, e.g., they can be M-ary symbols. There is no notion of distance between members of a finite field, except the so-called discrete topology: two elements may be equal or not equal. There is no notion of approximation. Hence in error-correction coding, there is no notion of approximate recovery. An element of x is either recoverable, or it is not.

[0010] One type of data for which approximation is useful is signal data, such as image, video, or audio data. Signal data, in contrast to symbol data, has elements that are members of an infinite field, such as the rational numbers, the real numbers, or the complex numbers. In a digital computer, these elements must be represented by finite precision numbers, such as 16-bit signed integers or 32-bit floating-point numbers. However there is a natural notion of distance between two such elements, e.g., the absolute value of their difference. As a consequence, any data vector x containing signal data (such as an image or a frame of audio) can be approximated by another data vector {circumflex over (x)}, where the degree of approximation can be measured by the Euclidean distance, a perceptual metric, or some other measure.

[0011] It is possible to apply error-correction codes to the transmission of signal data over erasure channels, by treating each signal element as a discrete M-ary symbol. However, this is inefficient. Signals need not be transmitted exactly, provided the approximation error between the signal and its reconstruction is small, on average. Applying a high-redundancy error-correction code can ensure this, but such a solution is expensive in terms of transmission capacity. An alternative is to use “unequal error protection,” which is a technique in which only the most important parts of the signal are protected with high redundancy. The less important parts are protected with low redundancy or not at all. One way to implement unequal error protection is the following. If the signal vector x contains elements that are 16-bit integers, then x can be stratified into 16 “bit-planes”. Then each bit-plane can be protected using a binary error-correction code with a redundancy commensurate with the importance of the bit-plane in approximating the signal vector x. (See for example J. Hagenauer, “Rate-Compatible Punctured Convolutional Codes (RCPC Codes) and their Applications,” IEEE Trans. on Communications, April 1988; A. Albanese, J. Blömer, J. Edmonds, M. Luby, and M. Sudan, “Priority Encoding Transmission,” IEEE Trans. Information Theory, Nov. 1996; G. Davis and J. Danskin, “Joint Source and Channel Coding for Image Transmission Over Lossy Packet Networks,” SPIE Conf on Wavelet Applic. of Digital Image Processing, August 1996.)

[0012] Even with the technique of unequal error protection, error-correction codes may not be ideally suited to protection of signal data. Ideally, all of the received bits should be used to improve the approximation of the reconstruction. However, in error-correction codes for erasure channels, if more than a critical number of symbols are received, then the additional symbols are ignored, while if fewer than the critical number of symbols are received, then any symbols not in the systematic portion of the code are ignored. In other words, if an error-correction code can correct a symbol exactly then it does not try any harder, while if it cannot correct a symbol exactly then it gives up. Although this may be ideal behavior for symbolic data, it is not ideal behavior for signal data. For signal data, the error control mechanism should try to approximate the transmitted signal with whatever information it has at its disposal.

[0013] Signal data is often voluminous. For example, a 512-by-512-pixel image has 262,144 pixel values. Therefore it is important that signal-processing techniques be computationally efficient.

[0014] Thus, there is a need for a method and apparatus that can efficiently reconstruct, to a suitable quality of approximation, an image or other set of signal data that has been transmitted on an erasure channel, or for some other reason has missing values.

SUMMARY OF THE INVENTION

[0015] The present invention provides improved communication of information, and in particular, a computerized method and/or system for communicating a set of data x having K dimensions. The invention enhances the recoverability of a set of data x having K dimensions for transmission on an erasure channel. One embodiment of the invention includes generating an expansion y of the set of data x using an N×K multiple description overcomplete expansion y=Fx, wherein N>K and wherein at least some of the last N−K rows of F are not duplicative of any of the first K rows. This embodiment also includes quantizing the coefficients of y to unequal resolution.

[0016] The overcomplete expansion provides redundant information. In some embodiments, the data are communicated on a channel or stored to and retrieved from a storage medium, and a version of the original set of data x is then reconstructed or restored from the received or retrieved data. Some embodiments provide reconstruction of the original set of data x using linear programming, linear reconstruction, and/or conventional POCS- (Projections Onto Convex Sets)-based algorithms.

[0017] Some embodiments of the present invention provide an efficient POCS-(Projections Onto Convex Sets)-based algorithm for consistent reconstruction from multiple descriptions of overcomplete expansions. In particular, image data can benefit from the reconstructions provided by the present invention, even when too little data is available for a perfect recreation of the original image. Similarly, audio information is another type of data that can benefit. The algorithm operates in the data space R^Krather than in the expanded space R^K, N>K. By constructing the frame from two (or more) complete transform bases, all projections can be expressed in terms of forward or inverse transforms. Since such transforms are usually efficient to compute, the present invention can perform the reconstruction much faster than with previous methods. Indeed, the method of one embodiment provides overcomplete frame expansions of an entire image and reconstruction of the image after transmission through a channel that loses some of the coefficients.

[0018] Aspects of the present invention provides a computerized method and/or system for reconstructing a set of data x having K dimensions. The method and system enhance the recoverability of a set of data x having K dimensions for transmission on an erasure channel.

[0019] In some embodiments, the unequal quantization resolution is based on a significance of data to its recovery.

[0020] In some embodiments, the set of data has a bit-significant representation, and the overcomplete expansion of the set of data provides coding most significant bits with a higher redundancy and least significant bits with a lower redundancy.

[0021] In some embodiments, the set of data is a source image having a bit-plane representation, and wherein coefficients of the overcomplete expansion of the source image provide coding the most significant bits with a higher redundancy and the least significant bits with a lower redundancy. In some such embodiments, the method further includes varying the quantization stepsize for the different coefficients. In some such embodiments, the method further includes receiving at least some of the quantized coefficients and decoding the received quantized coefficients using a method based on alternating projections onto convex sets in order to generate a recovered set of data.

[0022] One embodiment of the present invention provides a computer-readable medium having instructions stored thereon for causing a computer to perform one or more of the methods just described.

BRIEF DESCRIPTION OF THE DRAWINGS

[0023] FIG. 1 is a block diagram of a computer system 80 on which the present invention may be implemented.

[0024] FIG. 2A is a block diagram of a generic multiple-description system.

[0025] FIG. 2B is a block diagram of a quantizer-entropy encoder and a decoder.

[0026] FIG. 2C is a diagram of a quantization method used by the present invention.

[0027] FIG. 3A is a graphical representation of a reconstruction of a two-dimensional value 160 from a quantized three-dimensional description.

[0028] FIG. 3B is a graphical representation of a reconstruction of a two-dimensional value 178 from a quantized three-dimensional description.

[0029] FIG. 4 is a graphical representation of the sound-to-noise ratio (SNR) comparison of an MD POCS reconstruction according to the present invention versus that of an FEC system.

[0030] FIG. 5 is a recovered image of Lena that had one-eighth of the components missing (erasures).

[0031] FIG. 6 is a recovered image of Lena that had one-eighth of the components missing (erasures).

[0032] FIG. 7A is a flowchart 700A of one embodiment of the present invention.

[0033] FIG. 7B is a flowchart 700B of one embodiment of the present invention.

[0034] FIG. 7C is a flowchart 700C of one embodiment of the present invention.

[0035] FIG. 8A is a flowchart 800A of one embodiment of the present invention.

[0036] FIG. 8B is a flowchart 800B of one embodiment of the present invention.

[0037] FIG. 8C is a flowchart 800C of one embodiment of the present invention.

[0038] FIG. 9 is a block diagram of a circuit apparatus 900 that implements one embodiment of the present invention.

[0039] FIG. 10 is a flowchart 200 of one embodiment of the present invention.

[0040] FIG. 11 shows peak signal to noise ratio results for Lena showing reconstruction results for equal error protection.

[0041] FIG. 12 shows reconstruction results for unequal error protection.

[0042] FIG. 13 shows components of the vectors y₁=F₁x and y₂=F₂x depicting which bits are to be transmitted in one embodiment of unequal error protection

DETAILED DESCRIPTION

[0043] In the following detailed description of exemplary embodiments of the invention, reference is made to the accompanying drawings that form a part hereof, and in which is shown by way of illustration specific exemplary embodiments in which the invention may be practiced. These embodiments are described in sufficient detail to enable those skilled in the art to practice the invention, and it is to be understood that other embodiments may be utilized and that logical, mechanical, electrical and other changes may be made without departing from the spirit or scope of the present invention. The following detailed description is, therefore, not to be taken in a limiting sense, and the scope of the present invention is defined only by the appended claims.

[0044] The detailed description is divided into multiple sections. A first section describes the hardware and operating environment in the form of a computer system 80 that implements the current invention. This is followed by a description of examples of the initial transformation functions to be used, along with a mechanism for such transformations. Then a section is provided on the recovery transformations that take the received data from a channel and recover a version of the initial data. One particular embodiment is described in which an error-correction overcomplete expansion (or initial redundancy-adding transformation) is used to generate the redundant information to be transmitted, and an error-correction recovery iterative transformation process is used to restore the received data towards its original state. The factors involved in choosing the initial transformations and the recovery transformations are discussed, followed by a conclusion which describes some potential benefits and describes further alternative embodiments.

Hardware and Operating Environment

[0045] The invention is directed to a communications system. In one embodiment, the invention includes a hardware- and software-operating environment in which an encoder/transmitter operates in one computer, a receiver/decoder operates in another computer, and they are connected by a communications link. In some embodiments, the communications link is (or includes) a storage medium, and in some such embodiments the encoding and decoding computers are one and the same.

[0046] FIG. 1 provides a brief, general description of a suitable computing environment in which the either the encoder-transmitter or the receiver-decoder, or both, of the invention may be implemented. The invention will hereinafter be described in the general context of computer-executable program modules containing instructions executed by a personal computer (PC). This is one embodiment of many different computer configurations, some including specialized hardware circuits to enhance performance, that may be used to implement the present invention. Program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. Those skilled in the art will appreciate that the invention may be practiced with other computer-system configurations, including hand-held devices, multiprocessor systems, microprocessor-based programmable consumer electronics, network personal computers (“PCs”), minicomputers, mainframe computers, and the like which have multimedia capabilities. The invention may also be practiced in distributed computing environments where tasks are performed by remote processing devices linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote memory storage devices.

[0047] FIG. 1 shows a general-purpose computing or information-handling system 80. This embodiment includes a conventional personal computer (PC) 20, that includes processing unit 21, system memory 22, and system bus 23 that couples the system memory and other system components to processing unit 21. In other embodiments other configurations are used in PC 20. System bus 23 may be any of several types, including a memory bus or memory controller, a peripheral bus, and a local bus, and may use any of a variety of bus structures. In the embodiment shown, system memory 22 includes read-only memory (ROM) 24 and random-access memory (RAM) 25. A basic input/output system (BIOS) 26, stored in ROM 24, contains the basic routines that transfer information between components of personal computer 20. BIOS 26 also contains start-up routines for the system. Personal computer 20 further includes hard disk drive 27 having one or more magnetic hard disks onto which data is stored and retrieved for reading from and writing to hard-disk-drive interface 32, magnetic disk drive 28 for reading from and writing to a removable magnetic disk 29, and optical disk drive 30 for reading from and/or writing to a removable optical disk 31 such as a CD-ROM, DVD or other optical medium. Hard disk drive 27, magnetic disk drive 28, and optical disk drive 30 are connected to system bus 23 by a hard-disk drive interface 32, a magnetic-disk drive interface 33, and an optical-drive interface 34, respectively. The drives 27, 28, and 30 and their associated computer-readable media 29, 31 provide nonvolatile storage of computer-readable instructions, data structures, program modules and other data for personal computer 20. Although the exemplary environment described herein employs a non removable hard disk in a hard disk drive 27, a removable magnetic disk 29 and a removable optical disk 31, those skilled in the art will appreciate that other types of computer-readable media which can store data accessible by a computer may also be used in the exemplary operating environment. Such media may include magnetic tape cassettes, flash-memory cards, digital versatile disks (DVD), Bernoulli cartridges, RAMs, ROMs, and the like.

[0048] In various embodiments, program modules are stored on the hard disk drive 27, magnetic disk 29, optical disk 31, ROM 24 and/or RAM 25 and may be moved among these devices, e.g., from hard disk drive 27 to RAM 25. Program modules include operating system 35, one or more application programs 36, other program modules 37, and/or program data 38. A user may enter commands and information into personal computer 20 through input devices such as a keyboard 40 and a pointing device 42. Other input devices (not shown) for various embodiments include one or more devices selected from a microphone, joystick, game pad, satellite dish, scanner, and the like. These and other input devices are often connected to the processing unit 21 through a serial-port interface 46 coupled to system bus 23, but in other embodiments they are connected through other interfaces not shown in FIG. 1, such as a parallel port, a game port, or a universal serial bus (USB) interface. A monitor 47 or other display device also connects to system bus 23 via an interface such as a video adapter 48. In some embodiments, one or more speakers 57 or other audio output transducers are driven by sound adapter 56 connected to system bus 23. In some embodiments, in addition to the monitor 47, system 80 includes other peripheral output devices (not shown) such as one or more of scanner, printer, microphone, joystick, game pad, satellite dish, scanner, and the like.

[0049] In some embodiments, personal computer 20 operates in a networked environment using logical connections to one or more remote computers such as remote computer 49. Remote computer 49 may be another personal computer, a server, a router, a network PC, a peer device, or other common network node. Remote computer 49 typically includes many or all of the components described above in connection with personal computer 20; however, only a storage device 50 is illustrated in FIG. 1. The logical connections depicted in FIG. 1 include local-area network (LAN) 51 and a wide-area network (WAN) 52, both of which are shown connecting PC 20 to remote computer 49; typical embodiments would only include one or the other. Such networking environments are commonplace in offices, enterprise-wide computer networks, intranets and the Internet.

[0050] When placed in a LAN networking environment, PC 20 connects to local network 51 through a network interface or adapter 53. When used in a WAN networking environment such as the Internet, PC 20 typically includes modem 54 or other means for establishing communications over network 52. Modem 54 may be internal or external to PC 20, and connects to system bus 23 via serial-port interface 46 in the embodiment shown. In a networked environment, program modules, such as those comprising Microsoft® Word which are depicted as residing within PC 20 or portions thereof may be stored in remote-storage device 50. Of course, the network connections shown are illustrative, and other means of establishing a communications link between the computers may be substituted.

[0051] Software may be designed using many different methods, including object-oriented programming methods. C++ and Java are two examples of common object-oriented computer programming languages that provide functionality associated with object-oriented programming. Object-oriented programming methods provide a means to encapsulate data members (variables) and member functions (methods) that operate on that data into a single entity called a class. Object-oriented programming methods also provide a means to create new classes based on existing classes.

[0052] An object is an instance of a class. The data members of an object are attributes that are stored inside the computer memory, and the methods are executable computer code that act upon this data, along with potentially providing other services. The notion of an object is exploited in the present invention in that certain aspects of the invention are implemented as objects in some embodiments.

[0053] An interface is a group of related functions that are organized into a named unit. Some identifier may uniquely identify each interface. Interfaces have no instantiation; that is, an interface is a definition only without the executable code needed to implement the methods that are specified by the interface. An object may support an interface by providing executable code for the methods specified by the interface. The executable code supplied by the object must comply with the definitions specified by the interface. The object may also provide additional methods. Those skilled in the art will recognize that interfaces are not limited to use in or by an object-oriented programming environment.

Multiple-description Decoding of Overcomplete Expansions

[0054] Introduction

[0055] Multiple-description (“MD”) source coding is the problem of encoding a single source {Xi} into N separate binary descriptions at rates R₁, . . . , R_Nbits per symbol such that any subset S of the descriptions may be received and together decoded to an expected distortion D_Scommensurate with the total bit rate of S.

[0056] FIG. 2A shows a generic multiple-description system 200. Signal data x 201 enters the multiple-description encoder 203 on the left. The encoder (or sender) 203 produces N binary descriptions 223 of x at respective rates R₁, . . . , R_Nbits per source symbol (i.e., bits per element of x), and then transmits them over one or more separate channels 222. In some embodiments, binary descriptions 223 are embedded in packets (that include such additional information as routing, checking, destination and source information) for transmission on a packet network 222 (or storage to a storage medium).

[0057] Some subset (i.e., zero or more) of the channels 222 is broken, or equivalently, the corresponding descriptions 223 are erased. The remaining descriptions 223′ enter the decoder 204. The decoder (or receiver) 204 receives the remaining descriptions over the subset S of working channels 222, and reconstructs the signal {circumflex over (x)} 202 so as to approximate the input signal x with distortion D_Scommensurate with the total number of bits received. The signal {circumflex over (x)} 202 is called the “reconstruction” of x 201.

[0058] MD coding can be used in a variety of situations where sending signal data over a diversity of channels is helpful in overcoming the potential failure of some subset of channels. Some examples are the following.

[0059] Multiple descriptions can be transmitted to a mobile receiver from multiple antennae, so that when some antennae are hidden from the mobile receiver, the receiver is still able to reconstruct the signal with fidelity proportional to the number of visible antennae.

[0060] Multiple descriptions can be stored on multiple disks, so that if some of the disks crash or are taken off-line, the signal can still be reconstructed with fidelity proportional to the number of disks on-line.

[0061] Multiple descriptions can be transmitted over channels at different frequencies, so that when some channels are jammed or rendered useless by atmospheric fading, the receiver is still able to reconstruct the signal with adequate fidelity.

[0062] Multiple descriptions can be transmitted in different packets of a packet network, so that when some packets are lost due to congestion and buffer overflow in the network, the receiver is still able to reconstruct the signal with fidelity commensurate with the number of received packets.

[0063] Any erasure channel. All of the examples above, as well as the channel 222 of FIG. 2A, can be considered to be generalized erasure channels, and will herein be referred to simply as “erasure channels”.

[0064] What distinguishes MD coding, or MD quantization, from standard error-correction coding (for erasure channels) is that the data x 201 to be transmitted is treated as signal data rather than symbol data. That is, approximation is possible in order to reduce the bit rates R₁, . . . , R_N.

[0065] MD coding can be seen as a generalization of layered coding (also known as progressive, embedded, or scalable coding), in that in layered coding, the possible subsets of received channels are nested, i.e., contiguous channels 1 through K are received, where K varies between 0 and N. In this case, channel 1 is considered a base layer, and each subsequent layer enhances the previous layer. In contrast, in MD coding, any subset of channels can be received and used to reduce the distortion between x and {circumflex over (x)}.

[0066] Early papers on multiple-description coding are information theoretic and consider the problem of determining for N=2 the set of rates and expected distortions {(R₁, R₂, D₁, D₂, D_1,2)} that are asymptotically achievable, (see, e.g.,

[0067] 1 J. K. Wolf, A. D. Wyner, and J. Ziv. Source coding for multiple descriptions. Bell System Technical Journal,59(8):1417-1426, October 1980; herein “Wolf, et al.”;

[0068] 2 T. M. Cover and A. El Gamal. Achievable rates for multiple descriptions. IEEE Trans. Information Theory,28(6):851-857, November 1982; herein “Cover et al.”;

[0069] 3 L. H. Ozarow. On the source coding problem with two channels and three receivers. Bell System Technical Journal,59(8):1909-1922, October 1980; herein “Ozarow”;

[0070] 4 R. Ahlswede. The rate-distortion region of a binary source for multiple descriptions without excess rate. IEEE Trans. Information Theory,31:721-726, November 1985; herein “Ahlswede”;

[0071] 5 Z. Zhang and T. Berger. Multiple-description source coding with no excess marginal rate. IEEE Trans. Information Theory,41(2):349-357, March 1995; herein “Zhang et al.”).

[0072] More recent papers consider the problem of designing practical multiple-description quantizers, and their use over erasure channels.

[0073] The papers on practical MD quantization have so far taken three main and distinctly different approaches. In the first approach, pioneered by Vaishampayan, MD scalar, vector, or trellis quantizers are designed to produce N=2 descriptions, using a generalized Lloyd-like clustering algorithm that minimizes the Lagrangian of the rates and expected distortions R₁, R₂, D₁, D₂, D_1,2(see:

[0074] 6a V. A. Vaishampayan. “Vector quantizer design in diversity systems,” In Proc. Conf. On Information Sciences, Sep. 2^nd, 1991; herein “Vaishampayan 1”;

[0075] 6b V. A. Vaishampayan. “Design of multiple-description scalar quantizers,” IEEE Trans. Information Theory,39(3):821-834, May 1993; herein “Vaishampayan 2”;

[0076] 7 V. A. Vaishampayan. “Application of multiple-description codes to image and video transmission,” In Proc.7th Int'l Workshop on Packet Video, Brisbane, Australia, March 1996; herein “Vaishampayan 3”;

[0077] 8 S. D. Servetto, K. Ramchandran, V. Vaishampayan, and K. Nahrstedt. “Multiple description wavelet based image coding,” In Proc. Int'l Conf. Image Processing, Chicago, Ill., October 1998. IEEE, herein “Servetto et al.”;

[0078] 9 H. Jafarhkani and V. Tarokh. “Multiple description trellis coded quantization,” In Proc. Int'l Conf Image Processing, Chicago, Ill., October 1998. IEEE, herein “Jafarhkani et al.”;

[0079] 10 M. Srinivasan and R. Chellappa. “Multiple description subband coding,” In Proc. Int'l Conf. Image Processing, Chicago, Ill., October 1998. IEEE; herein “Srinivasan”.)

[0080] In the second approach, pioneered by Wang, Orchard, and Reibman, MD quantizers are constructed by separately describing (i.e., quantizing and coding) the N coefficients of an N×N block linear transform, which has been designed to introduce a controlled amount of correlation between the transform coefficients (see:

[0081] 11 Y. Wang, M. T. Orchard, and A. R. Reibman. Multiple description image coding for noisy channels by pairing transform coefficients. In Proc. Workshop on Multimedia Signal Processing, pages 419-424. IEEE, Princeton, N.J., June 1997; herein “Wang et al. 1”,

[0082] 12 M. T. Orchard, Y. Wang, V. A. Vaishampayan, and A. R. Reibman. Redundancy rate-distortion analysis of multiple description coding. In Proc. Int'l Conf Image Processing, Santa Barbara, Calif., October 1997. IEEE; herein “Orchard et al.”,

[0083] 13 V. K. Goyal and J. Kovaćević. Optimal multiple description transform coding of Gaussian vectors. In Proc. Data Compression Conference, pages 388-397, Snowbird, Utah, March 1998. IEEE Computer Society; herein “Goyal et al. 1”,

[0084] 14 V. K. Goyal, J. Kovaćevi{haeck over (c)}, R. Arean, and M. Vetterli. Multiple description transform coding of images. In Proc. Int'l Conf. Image Processing, Chicago, Ill., October 1998. IEEE; herein “Goyal et al. 2”,

[0085] 15 Y. Wang, M. T. Orchard, and A. R. Reibman. Optimal pairwise correlating transforms for multiple description coding. In Proc. Int'l Conf Image Processing, Chicago, Ill., October 1998. IEEE; herein “Wang et al. 2”.)

[0086] In the third approach, pioneered by Goyal, Kovaćević, and Vetterli, MD quantizers are constructed by separately describing the N coefficients of an overcomplete N×K tight frame expansion (see Goyal et al. 2 cited above and

[0087] 16 V. K. Goyal, J. Kovaćević, and M. Vetterli. Multiple description transform coding: robustness to erasures using tight frame expansions. In Proc. Int'l Symp. Information Theory, page 408, Cambridge, Mass., August 1998. IEEE; herein “Goyal et al. 3”.).

[0088] The present invention contributes to this last category. For completeness, it should be mentioned that a number of other papers take yet a fourth approach, in which the natural correlation between symbols is exploited for reconstruction. For example, odd pixels can be predicted from even pixels, and vice versa. This fourth approach is similar to the second approach above, except that the transform is not actively designed. This fourth approach is considered by the inventors to be more closely related to standard error-resilience techniques.

[0089] In MD quantization using overcomplete (frame) expansions, a K-dimensional input signal x is represented by an N-dimensional vector y=Fx, where N>K. F 210 is the so-called frame operator, whose N rows span the K-dimensional signal space. This is the same setup as in error-correction coding, except that in MD coding, the elements of x and y (as well the elements of F) are members of an infinite field (such as the rational numbers, the real numbers, or the complex numbers) rather than being members of a finite field. Without loss of generality, the elements of x, y, and F are considered to be real numbers for some embodiments of the present invention.

[0090] Herein, the term “overcomplete expansion” and more specifically the term “multiple description overcomplete expansion” are used to designate the signal expansion y=Fx used in MD coding, in contrast to the analogous expansion y=Fx over a finite field used in error-correction coding.

[0091] Because the elements of y 217 can have an infinite number of possible values, y must first be quantized and entropy coded into a finite number of bits before being transmitted or stored. One such embodiment is shown in FIG. 2B.

[0092] In this embodiment, the N-dimensional vector y 217 is scalar quantized by quantizer 218. In scalar quantization, each element y_iof the N-dimensional vector y is mapped (i.e., using quantizer mapping 288) by a quantizer Q_ito the index j_iof the quantization interval [l_i,u_i) which contains y_i. For example, quantizer Q_imay map y_ito 0 if y_iis between −5 and 5, to 1 if y_iis between 5 and 15, to 2 if y_iis between 15 and 25, and so forth, as illustrated in FIG. 2C. The output of quantizer 218 is the set of indices j 221. The inverse quantizer mapping 296 (performed by block 226) provides the received values 227, that include one or more of l_i,u_i,ŷ_i.

[0093] Thus, if y_iis the value 27.109, j_iwould be the index 3, and the interval [l_i,u_i) would be [25,35). Often y_iis represented by the midpoint or centroid of the interval, which would be ŷ_i=30 in this case. If the interval lengths are all constant, then the quantizer is called a uniform scalar quantizer, and the interval length is called the quantizer stepsize. This invention does not require the interval lengths to be constant, either within a quantizer or between quantizers. Thus, in other embodiments, non-uniform interval lengths are used. In particular, in one embodiment (explained further in a section below), the interval length changes between quantizers, to provide unequal redundancy encoding, based on data significance. In principle, it is also possible to perform vector quantization on the N-dimensional vectory. (See A. Gersho and R. M. Gray, Vector Quantization and Signal Compression, Kluwer, 1991, herein Gersho and Gray.) However only scalar quantization is discussed herein.

[0094] In the embodiment shown in FIG. 2B, the indices j_iare grouped, and each group of indices is independently entropy encoded by entropy encoder 219. Entropy encoding is the process of mapping a set of discrete symbols into a binary string whose length is approximately equal to the entropy of the set of symbols. Herein the term “entropy encoding” is used generally to denote the process of creating a binary string, or “binary description” 223 from a group of indices. This can be done with a variety of well-known techniques (see for example Gersho and Gray; A. Said and W. A. Pearlman, “A New, Fast, and Efficient Image Codec Based On Set Partitioning In Hierarchical Trees,” IEEE Trans. Circuits and Systems for Video Technology, June 1996; E. Malvar, “Fast Progressive Wavelet Coding,” Data Compression Conference, Snowbird, Utah, 1999.). The number of descriptions 223 emanating from the MD encoder is equal to the number of groups. Since groups may contain only a single index j_i, it is possible for there to be up to N descriptions 223.

[0095] In some embodiments, each of the binary descriptions 223 is embedded into a packet that is transmitted on a packet network or channel 222. Some of the descriptions 223 may be erased in the transmission or storage channel 222. In that case, only the indices in groups whose descriptions are received 223′ can be entropy decoded by entropy decoder 224. Entropy decoding is the inverse of entropy encoding, i.e., the process of mapping a binary string back into a set of discrete symbols. The number of indices that are entropy decoded is only N′≦N after potential erasures. Each decoded index j_i225 is an indicator of the fact that the corresponding coefficient y lies in a particular interval [l_i,u_i). The inverse quantizer Q_i⁻¹226 therefore maps j_iinto the tuple [l_i,u_i), possibly along with the reconstruction level ŷ_i(together denoted as received information 227), as illustrated in FIG. 2B. The reconstruction 290 of the vector {circumflex over (x)} from the received information l_i,u_i,ŷ_i227 can take a number of forms, described next.

[0096] Without loss of generality, assume that descriptions of the first N′ coefficients are received, and that descriptions of the last N″=N−N′ coefficients are erased. Let y′ denote the first N′ coefficients, and let y″ denote the last N” coefficients, so that

y = [\begin{matrix} y^{'} \\ y^{″} \end{matrix}] = [\begin{matrix} F^{'} \\ F^{″} \end{matrix}] x = F x,

[0097] where F′ is N′×K and F″ is N″×K. A classical way for the decoder to reconstruct x from the received quantized coefficients ŷ′ is to use the linear reconstruction

[0098] {circumflex over (x)}_lin=(F′)⁺ŷ′ (Equation 1)

[0099] where (F′)⁺ is the pseudo-inverse of F′. The pseudo-inverse can be computed from the singular value decomposition F′=U·diag (σ₁, . . . , σ_N′)V¹(see 17 B. Noble and J. W. Daniel. Applied Linear Algebra. Prentice-Hall, Englewood Cliffs, N.J., 2nd edition, 1997.)

[0100] as (F′)=V·diag (σ₁⁻¹, . . . , σ_N′⁻¹)·U¹. This is equal to (F′)⁺=((F)^t(F′))⁻¹(F)^twhen F′ has full rank, i.e., when at least K descriptions are received (assuming any K rows of F are linearly independent). It can be shown that the reconstruction (Equation 1) has the property that

\begin{matrix} {\hat{x}}_{lin} = \arg \min_{x} {‖ {\hat{y}}^{'} - F^{'} x ‖}^{2} . & (Equation 2) \end{matrix}

[0101] That is, when N′≧K, it chooses {circumflex over (x)}_linεR^Kto be the coordinates of the point F′{circumflex over (x)}_linin the K-dimensional subspace spanned by the columns of F′ which is closest to ŷεR^N′, i.e., it projects ŷ′ onto the subspace F′R^N′. Furthermore, when N′<K, i.e., when the x that minimizes (Equation 2) is not unique, then the reconstruction (Equation 1) chooses an x with minimum norm.

[0102] The linear reconstruction (Equation 2) is not statistically optimal. The optimal reconstruction {circumflex over (x)}_opt, which minimizes the expected (squared error) distortion E∥X−{circumflex over (X)}_opt∥², is the conditional mean of X given the descriptions y_iε[l_i, u_i) received. That is,

[0103] {circumflex over (x)}_opt=E[X|Q(F′X)=ŷ′]=E[X|l′≦F′X<u′|, (Equation 3)

[0104] where the relations l′≦y′ and y′<u′ are to be taken componentwise. Unfortunately, the conditional expected value of X given that it lies in the region Q⁻¹(ŷ′)={x:l′≦F′x<u′} is hard to compute. FIG. 3A shows for K=2 and N′32 3 a region Q⁻¹(ŷ′) 160 for some ŷ′. Note that the regions Q⁻¹(ŷ′) for different ŷ′ are all dissimilar in general.

[0105] FIG. 3A is a graphical representation 167 of the set 160 of all possible reconstructions consistent in a quantized three-dimensional description. The three-dimensional description includes descriptions 163 having a quantization interval 164 with a lower limit l₁and upper limit u₁, description j₂161 having a quantization interval 162 with a lower limit l₂and upper limit u₂, and description j₃165 having a quantization interval 166 with a lower limit l₃and upper limit u₃. The reconstructed value is somewhere within shaded area 160. Because of the effects of quantization on each of the descriptions, an area at the intersection of the descriptions remains. If only two descriptions were used, the intersection would be a square. When three descriptions are used, the area is reduced to quadrilateral 160.

[0106] Although the optimal reconstruction (Equation 3) is difficult to compute, one thing is certain: x lies in Q⁻¹(ŷ′), and hence, since Q⁻¹(ŷ′) is convex, the optimal reconstruction (Equation 3) lies in Q⁻¹(ŷ′) Any reconstruction {circumflex over (x)} which does not lie in Q⁻¹(ŷ′) is said to be inconsistent (See V. K. Goyal, M. Vetterli, and N. T. Thao. Quantized overcomplete expansions in Rⁿ: Analysis, synthesis, and algorithms. IEEE Trans. Information Theory,44(1):16-31, January 1998; herein “Goyal et al. 4”). Goyal et al. 4 describes one conventional POCS reconstruction method. FIG. 3, adapted from Goyal et al., shows again for K=2 and N′=3 an inconsistent reconstruction from the linear projection (Equation 1). It is intuitive that consistent reconstructions have smaller expected squared error distortion than inconsistent reconstructions. In fact, Goyal et al. show that while the expected distortion from linear reconstructions is asymptotically proportional to 1/N, the expected distortion from consistent reconstructions is O(1/N²) in at least one DFT-based case, and they conjecture this to be true under very general conditions, when the frame is tight.

[0107] FIG. 3B is a graphical representation 168 of an inconsistent reconstruction 177 from a quantized three-dimensional description. The centroid ŷ′ 187 of cube 180 would project to {circumflex over (x)}_lin177, which is a distance 176 in the subspace away from a consistent reconstruction {circumflex over (x)}_con178 (which may be thought of as in the intersection of cube 180 with plane F′R^K170). Centroid ŷ′ 187 is a distance 186 out of the subspace away from {circumflex over (x)}_lin177.

[0108] An algorithm used by some embodiments of the present invention for producing consistent reconstructions {circumflex over (x)}_conεQ⁻¹(ŷ′) is the POCS (Projections Onto Convex Sets) algorithm (see

[0109] 19 D. C. Youla. Mathematical theory of image restoration by the method of convex projections. In H. Stark, editor, Image Recovery: Theory and Applications. Academic Press, 1987; herein “Youla”.).

[0110] In the POCS algorithm for M=2 (i.e., where only two functions are used, as more fully described below), an arbitrary initial point p_tεRⁿis alternately projected onto two convex sets Q⊂Rⁿand P⊂Rⁿ,

\begin{matrix} q_{t + 1} = \arg \min_{q \in Q} {‖ p_{t} - q ‖}^{2} & (Equation 4) \\ p_{t + 1} = \arg \min_{p \in P} {‖ q_{t + 1} - p ‖}^{2} & (Equation 5) \end{matrix}

[0111] until p_tand q_tconverge to the intersection of P and Q, or if the intersection is empty, until p_tand q_trespectively converge to the sets {pεP:∥p−q∥²≦∥p′−q∥², ∀p′εP, ∀qεQ} and {qεQ:∥q−p∥²≦∥q′−p∥², ∀q′εQ, ∀pεP}. The POCS algorithm described above can be extended in the obvious way to more than two sets. For example, a point in the intersection of three convex sets P, Q, and R is given by the limit of

r_{t + 1} = \arg \min_{r \in R} {‖ p_{t} - r ‖}^{2}, q_{t + 1} = \arg \min_{q \in Q} {‖ r_{t + 1} - q ‖}^{2}, and

p_{t + 1} = \arg \min_{p \in P} {‖ q_{t + 1} - p ‖}^{2} .

[0112] For consistent reconstruction of xεR^Kfrom the quantized frame expansion ŷ′=Q(F′x)εR^N′, Goyal et al. (Goyal et al. 4) suggest alternately projecting the initial point ŷ′ onto the two convex sets Q=F′R^KεR^N′ and P=Q⁻¹(ŷ′)εR^N′, as shown in FIG. 3. The first projection (Equation 4) onto the linear subspace F′R^Kcan be accomplished, as usual, by the pseudo-inverse (Equation 1). The second projection (Equation 5) onto the quantization bin Q⁻¹(ŷ′) can be accomplished by component-wise clipping to the quantization bin. That is, the vector ŷ′, l′≦ŷ′≦u′, closest to an arbitrary vector ŷ′εR^N′ is given component-wise by

[0113] {tilde over (y)}_i′=clip(y_i′, l_i, u_i)=min {max{y_{i′, l}_i}, u_i}. (Equation 6)

[0114] Unfortunately, computation of the pseudo-inverse (F′)⁺ requires O(K²N′) operations; the projections themselves require O(KN′) operations each.

[0115] To reduce the computational complexity of consistent reconstruction, Goyal, Vetterli, and Thao instead suggest finding {circumflex over (x)}εQ⁻¹(ŷ′) by solving the linear program max c^t{circumflex over (x)} subject to

[\begin{matrix} F^{'} \\ - F^{'} \end{matrix}] \hat{x} \leq [\begin{matrix} u \\ - l \end{matrix}],

[0116] for an arbitrary objective functional c. They furthermore suggest that by varying c, it may be possible to find all the vertices of Q⁻¹(ŷ′), whereupon they can be averaged to approximate the region's centroid. (However, they do not appear to follow this latter suggestion.) Although this method avoids the high cost of computing the pseudo-inverse, the complexity of the simplex algorithm for solving the linear program is still O(KN) operations per pivot. This complexity does not present a problem when K and N are small. Goyal, Kova{haeck over (c)}ević, and Vetterli (see Goyal 3 and Goyal 2) apparently apply the algorithm for K=8 and N′ up to 10.

[0117] We are more interested in decoding multiple descriptions of overcomplete expansions based on overlapping functions, such as provided by wavelet, subband, or lapped transforms. In this case, the N×K frame operator F typically operates on an entire image at a time, for which K=512×512=262,144 is common; N may be twice that. Clearly, in such cases it is not feasible to require O(KN) operations for consistent reconstruction. For wavelet, subband, or lapped transforms, where N=K, F is sparse. In this case, consistent reconstruction can be performed using O(KL) operations, where L is the length of the support of the basis functions. Since L is usually on the order of a few hundred, such reconstruction is eminently feasible, and is used in all modern subband decoders. Both the pseudo-inverse and pivot operations destroy the sparsity of F. The present invention provides an algorithm for consistent reconstruction from multiple descriptions of overcomplete expansions that preserves the efficiency of the sparse representation of F when the basis functions have finite support L. That is, the algorithm has complexity O(KL). The algorithm of the present invention is based on POCS, but the projections are performed in the lower dimensional space R^K, rather than in the space R^N′.

[0118] In some embodiments (further explained in a section below), the encoding transform is selected to provide unequal redundancy encoding, based on data significance. In one such embodiment, the starting set of data is a source image having a bit-plane representation, and wherein coefficients of the overcomplete expansion of the source image provide coding the most significant bits with a higher redundancy and the least significant bits with a lower redundancy.

[0119] Decoding Algorithm

[0120] Let F₁, . . . , F_Mbe M invertible K×K transforms. For example, F₁may be a wavelet transform (over an image xεR^Ksuitably extended), F₂may be the identity transform, another wavelet transform, or the same wavelet transform over a 1-pixel shift of the image and so forth for F₃, . . . , F_M. Note: in some embodiments as described below in FIG. 7A and FIG. 8A, M=2, and only F₁and F₂are used. However, in general, M can be any integer 2 or larger, as described below in FIG. 7B and FIG. 8B. Let the K-dimensional vectors y₁=F₁x, . . . , y_M=F_Mx be the M corresponding sets of transform coefficients for xεR^K. Then

\begin{matrix} y = [\begin{matrix} y_{l} \\ ⋮ \\ y_{M} \end{matrix}] = [\begin{matrix} F_{l} \\ ⋮ \\ F_{M} \end{matrix}] x = F x & (Equation 7) \end{matrix}

[0121] defines an overcomplete N×K frame expansion of x with redundancy N/K=M. The expansion is tight if F₁, . . . , F_Mare orthonormal. (A frame expansion y=Fx is “tight” if there exists a positive constant M such that for all x, ∥Fx∥²=M∥x∥². In general, for the multiple description scenario, tightness of the original frame F is of little consequence, because the received frame F′ will, in general, not be tight.)

[0122] The present invention does not require orthonormality of F₁, . . . , F_M, but it is best if F₁, . . . , F_Mare orthonormally related, i.e., for all i,j, F_ij=F_jF_i⁻¹is orthonormal, at least approximately, to ensure convergence of the POCS algorithm.

[0123] Let ŷ₁, . . , ŷ₂be the quantized versions of y₁, . . . , y_M, respectively, such that ŷ₁, . . . ŷ_Mlie (componentwise) between upper and lower quantization cell boundary vectors l₁≦ŷ_i≦u₁through l_M≦ŷ_M≦u_M, respectively.

[0124] Let R₁∈{1, . . . , K} be the set of indices of the descriptions received by the decoder for ŷ₁, let R₂∈{1, . . . , K} be the set of indices of the descriptions received by the decoder for ŷ₂and so forth through R_M. Descriptions not received by the decoder include those that have been erased as well as not sent at all.

[0125] A reconstruction {circumflex over (x)} is consistent with the received descriptions if and only if it lies in the intersection of the following M sets (where M is a number 2 or larger and a similar equation is provided for all intermediate values between 1 and M):

[0126] P₁={x:l_1,i≦(F₁x)_i≦u_1,i, iεR₁} (Equation 8)

[0127] .

[0128] .

[0129] .

[0130] P_M={x:l_M,i≦(F_Mx)_i≦u_M,i, iεR_M} (equation 9)

[0131] The following steps illustrate an exemplary basic algorithm of one embodiment of the present invention for finding a consistent reconstruction of x from the received descriptions.

[0132] 1. Initialization. Start from an initial point in F₁P₁:

\begin{matrix} p_{1, i} (0) = {\hat{y}}_{1, i} & if i \in R_{1} & (i . e ., received descriptions) and \\ p_{1, i} (0) = 0 & if i \notin R_{1}; & (i . e ., descriptions not received) \end{matrix}

[0133] 2. For each m=1, . . . , M−1 in turn

[0134] a. Transform p_m(t) into the coordinate system of F_m+1:

[0135] {tilde over (p)}_m+1(t)=F_m+1F_m⁻¹P_m(t)=F_m,m−1p_m(t);

[0136] b. Project {tilde over (p)}_m+1(t) onto F_m+1P_m+1:

\begin{matrix} p_{m + 1, i} (t) = \min {\max {{\tilde{p}}_{m + 1, i} (t), l_{m + 1, i}} u_{m + 1, i}} & if i \in R_{2} \\ p_{m + 1, i} (t) = {\tilde{p}}_{m + 1, i} (t) & if i \notin R_{2} \end{matrix}

[0137] 3. Transform p_M(t) into the coordinate system of F₁:

[0138] {tilde over (p)}₁(t+1)=F₁F_M⁻¹p_M(t)=F_M,1p_M(t)

[0139] 4. Project {tilde over (p)}₁(t+1) onto F₁P₁:

\begin{matrix} p_{1, i} (t + 1) = \min {\max {{\tilde{p}}_{1, i} (t + 1), l_{1, i}}, u_{1, i}} & if i \in R_{1} \\ p_{1, i} (t + 1) = {\tilde{p}}_{1, i} (t + 1) & if i \notin R_{1} \end{matrix}

[0140] 5. Checkfor convergence. If ∥p₁(t+1)−p₁(t)∥²> epsilon, then set t←t+1 and go to Step 2.

[0141] 6. Reconstruct x.

[0142] {circumflex over (x)}=F₁⁻¹p₁(t+1).

[0143] A reasonable value for epsilon is K, so that the squared error per pixel is within one gray level.

[0144] In other embodiments, convergence is checked at other points between 1 and M, and steps 5 and 6 are replaced by the following 5′ and 6′:

[0145] 5′. Check for convergence. If ∥p_n(t+1)−p_n(t)∥²> epsilon, then continue transforming and projecting until M, then set t←t+1 and go to Step 2, else:

[0146] 6′. Reconstruct x.

[0147] {circumflex over (x)}=F_n⁻¹p_n(t+1).

[0148] In other embodiments, step 5 (or 5′) above is replaced by a loop counter, and a predetermined number of iterations is performed (e.g., twenty iterations has been found to be sufficient to achieve sufficient convergence for certain embodiments and certain channel-loss rates). In other embodiments, the number of loop iterations needed to achieve sufficient convergence is related to the number of missing coefficients in the received data ŷ, and the loop counter is set to a value based on the number missing values. Such embodiments avoid actually measuring convergence (which may take a significant amount of time), and instead use an empirically derived number of iterations that are expected to provide convergence.

[0149] In other embodiments, the initial point, the order of the projections onto the convex sets and other details inconsequential to the ultimate convergence of the algorithm may be different. However, it is understood that such variations do not depart from the spirit and scope of the present invention.

[0150] FIG. 7A is a flowchart 700A of one embodiment of the present invention, where M=2. FIG. 7B is a flowchart 700B of another, more generalized embodiment of the present invention, where M is any number 2 or larger. FIG. 7C is a flowchart 700C of another, more generalized embodiment of the present invention, where convergence checking is performed in the middle of the loop. Figures 7B and 7C are otherwise identical to FIG. 7A, which will now be described. Flowchart 700A represents an embodiment used to reconstruct a set of data {circumflex over (x)} from a received set of data y′, wherein M=2 and the received data are mapped back and forth between F₁P and F₂Q. Block 710 performs an initialization operation to initialize a point (e.g., a vector in N dimensional space) in F₁P, and to set t=0. In one such embodiment, the operations set forth in Step 1 above are performed in block 710. Control then passes to block 720. Block 720 performs an operation transforming p₁(t) into the coordinate system of F₂. In one such embodiment, the operations set forth in Step 1 above are performed in block 720. Control then passes to block 730. Block 730 performs an operation projecting p₂(t) onto F₂Q. In one such embodiment, the operations set forth in Step 3 above are performed in block 730. In one embodiment, the functions of blocks 720 and 730 are performed as a single function, labeled block 725.

[0151] Control then passes to block 740. Block 740 performs an operation transforming q₂(t+1) into the coordinate system of F₁. In one such embodiment, the operations set forth in Step 4 above are performed in block 740. Control then passes to block 750. Block 750 performs an operation projecting q₁(t+1) onto F₁P. In one such embodiment, the operations set forth in Step 5 above are performed in block 750. In one embodiment, the functions of blocks 740 and 750 are performed as a single function, labeled block 745.

[0152] Control then passes to block 760. Block 760 performs an operation checking for convergence and if not sufficiently converged, then iterating blocks 720 through 760. In one such embodiment, the operations set forth in Step 6 above are performed in block 760. Control then conditionally passes to block 770 if convergence has been achieved, otherwise control passes back to block 720. Block 770 performs an operation reconstructing {circumflex over (x)}=F₁⁻¹p₁(t+1) in order to generate a recovered set of data. In one embodiment, F₁and F₂are each K×K transforms. In some embodiments, F₁and F₂are each K×K linear transforms.

[0153] In another embodiment shown in FIG. 7B, the convergence check of block 760 is performed between block 730 and block 740, with control passing conditionally to block 770 if convergence has been achieved, otherwise control passes to block 740. In this embodiment, block 770 performs an operation reconstructing {circumflex over (x)}=F₂⁻¹q₂(t+1) in order to generate a recovered set of data.

[0154] In one embodiment, the received set of data y′ are the coefficients received from an erasure channel that transmitted an overcomplete expansion of an original image x, and the recovered set of data {circumflex over (x)} represents a reconstruction of the original image. In one such embodiment, the original image x is a portion of a larger image, for example, a 512-by-512-pixel portion of a 1024-by-1024 pixel image. In another embodiment, the original image x is a portion (e.g., one frame) of a sequence of images representing a motion picture. In another embodiment, the received set of data y′ are the coefficients received from an erasure channel that transmitted an overcomplete expansion of an original audio segment x, and the recovered set of data {circumflex over (x)} represents a reconstruction of the original audio segment.

[0155] The present invention uses any number of K×K transform functions and mappings (up to M transform-and-mapping operations) to reconstruct the received data, with typical embodiments shown in FIG. 8A, FIG. 8B, and FIG. 8C. In some embodiments, the original encoding performs an overcomplete N×K frame expansion using at least portions of each of the M transforms, as described below.

[0156] FIG. 8A is a flowchart 800A of one embodiment of the present invention for the case M=2. FIG. 8B is a flowchart 800B of one embodiment of the present invention for the general case M, where the convergence test 870 is performed after a complete set of transform-mappings. FIG. 8C is a flowchart 800C of one embodiment of the present invention for the general case M, where the convergence test 870 is performed at an arbitrary point in the transform-mappings. The term “Figure 8” refers generally to FIG. 8A, FIG. 8B, and FIG. 8C.

[0157] Flowchart 800A represents one embodiment used to generate an overcomplete expansion y from an initial set of data x, to transmit y through a lossy channel, and then to reconstruct a set of data {circumflex over (x)} from a received set of data y′. Block 201 represents a set of data x that is provided. In one embodiment, set of data x is an image, for example, a 512-by-512-pixel image (which, in some embodiments, is a portion of a large image) that is to be transmitted through erasure channel 222. Block 812 performs an overcomplete N×K frame expansion such as defined in Equation 7 above to generate set of data y 217. In one such embodiment, this is accomplished by using K×K linear transform F₁811 to generate a first set of data y₁815 and K×K linear transform F₂812 to generate a second set of data y₂816. Some or all of the first set of data y₁815 and some or all of the second set of data y₂816 are concatenated to form the set of data y 217. In some embodiments, block 210 implements a N×K transform F, wherein all of the first set of data y₁815 is included in y, and N−K coefficients of the second set of data y₂816 are included. In one such embodiment, the N−K coefficients of the second set of data y₂816 implement unequal error protection as further described in a section below. In some embodiments, this provides better reconstruction of the most important aspects of x, for example, the most-significant bits of pixels of an image.

[0158] In some embodiments, the rectangular blocks of FIG. 8 (e.g., 210, 830, 880) are implemented in storage area 36 of FIG. 1 (i.e., application programs), and the rounded blocks (e.g., 201, 227, 865) are implemented in storage area 38 of FIG. 1 (i.e., program data).

[0159] The set of data y 217 is then transmitted through erasure channel 222 and received by receive block 220. In some embodiments, channel 222 is a packet network having one or more parallel paths, some of which may be more lossy than others. That is, in some embodiments, channel 222 is implemented as a plurality of parallel channels, some of which may transmit more data than others, and some of which may lose more packets than others. In one such embodiment, channel 222 will fail to transmit some of the packets (generally a random subset will be lost), and thus set of data y′ has less than or equal to the corresponding coefficients of the set of data y, and the indices of the missing values of y′ are known.

[0160] In some embodiments, as shown in FIG. 8B, the set of data y 217 is quantized in quantizer block 218 and entropy encoded in entropy encoder block 219, both of which are implemented in transmitter block 214, and the received binary descriptions 223′ are entropy decoded by entropy decoder 224 and inverse quantized by inverse quantizer 226 both of receiver 220 (described further in the FIG. 2B description above).

[0161] In other embodiments, channel 222 is implemented as a noisy channel, in which noise will be detected in some of the received coefficients. In some such embodiments, those coefficients will be erased, and receiver 220 will mark the corresponding indices to indicate missing data, thus producing the equivalent of an erasure channel as described above, to output y′ 227.

[0162] In some embodiments, channel 220 will include a storage component, for example, a digital versatile disk (DVD), onto which a series of images will be recorded (e.g., a video movie). That is, channel 222 is the process of writing and reading to the storage medium, along with any losses that occur to data on the medium over time, and losses occurring in the read process. In some such embodiments, the series of images will be recorded with conventional error recovery at a low level, and with the overcomplete expansion of the present invention at a higher level. Then, when the movie is read from the DVD, and the low-level error recover cannot recover a given packet, that packet will be indicated as “erased” or lost, and presented as such in the received set of data y′ 227.

[0163] In the embodiment shown, the received set of data 227 is used as input data to the iterative process that includes 830, 840, 850, 860, 870, and 880. The recovery portion of method 800 starts by generating an initial point p₁(0) 829 in F₁P₁using the following equations:

p_{1, i} (0) = {\hat{y}}_{1, i} if i \in R_{1} and (i . e ., description is received) and

p_{1, i} (0) = 0 if i \notin R_{1}; (i . e ., description is not received)

[0164] and set t=0.

[0165] Block 830 operates to transform p₁(t) (829 or 865) into the coordinate system of F₂, effectively using the inverse transform of F₁and then the F₂transform. In some embodiments, this is accomplished using a single K×K transform F₁₂as follows:

[0166] {tilde over (p)}₂(t)=F₂(F₁⁻¹p₁(t))=F₁₂p₁(t).

[0167] Block 840 operates to project p₂(t) 835 onto F₂P₂as follows:

p_{2, i} (t) = \min {\max {{\tilde{p}}_{2, i} (t), l_{2, i}}, u_{2, i}} if i \in R_{2} and

p_{2, i} (t) = {\tilde{p}}_{2, i} (t) if i \notin R_{2} .

[0168] Block 850 operates to transform p₂(t) 845 into the coordinate system of F₁as follows:

[0169] {tilde over (p)}₁(t+1)=F₁(F₂⁻¹p₂(t))=F₂₁p₂(t).

[0170] Block 860 operates to project {tilde over (p)}₁(t+1) 855 onto F₁P₁as follows:

p_{1, i} (t + 1) = \min {\max {{\tilde{p}}_{1, i} (t + 1), l_{1, i}}, u_{1, i}} if i \in R_{1} and

p_{1, i} (t + 1) = p_{2, i} (t) if i \notin R_{1} .

[0171] In one embodiment, the “previous” version of p₁(t) is kept in storage of block 829, and the “current” version p₁(t+1) is kept in block 865, and blocks 829 and 865 are compared by block 870.

[0172] In FIG. 8B (which is otherwise similar to FIG. 8A), M can be other values besides 2. The initial overcomplete expansion F uses portions or all of M different functions F₁through F_M. FIG. 8B also shows the quantization and entropy encode operations of transmitter 214 and entropy decode and inverse quantization operations of receiver 220, as described for FIG. 2B above (although not shown in FIG. 8A, some embodiments of the system of FIG. 8A also include such functions). Successive transforms and projections are performed between blocks 845 and 854. Block 854 operates to transform the resulting p_M(t) into the coordinate system of F₁as follows: {tilde over (p)}₁(t+1)=F₁(F_M⁻¹p_M(t+1))=F_M1p_M(t+1).

[0173] For each of FIGS. 8A and 8B, block 870 operates to check for convergence as follows:

[0174] If ∥p₁(t+1)−p₁(t)∥²>ε, then set t←t+1 and go to block 830 (and, in one embodiment, load p₁(t+1) from block 865 into block 829 as the next p₁(t)). Block 880 operates to reconstruct x into the output block {circumflex over (x)} 202 as follows:

[0175] {circumflex over (x)}=F₁⁻¹p₁(t+1).

[0176] In FIG. 8C (which is otherwise similar to FIG. 8B), convergence checking is performed within the transformation-projection loop. In this embodiment, block 870 operates to check for convergence as follows: After the F_n−1,ntransform-projection of blocks 864, 865, 866, and 867, If ∥p_n(t+1)−p_n(t)∥²>ε, then continue, else block 880 operates to reconstruct x into the output block {circumflex over (x)} 202 as follows: {circumflex over (x)}=F_n⁻¹p_n(t+1). After the F_M1transform-projection of blocks 894, 895, 896, and 897, set t←t+1 and go to block 830 (and, in one embodiment, load p₁(t+1) from block 865 into block 829 as the next p₁(t)).

[0177] In some embodiments, the flowchart 800 (i.e., 800A, 800B or 800C) of FIG. 8 is implemented as a software-implemented function of system 80 of FIG. 1, thus making system 80 a special-purpose machine to encode and/or decode data that goes through a lossy channel 222. In various embodiments, channel 222 includes one or more of the following: hard disk 27, removable disk 29, optical disk 31, WAN 52, and/or LAN 51. In some embodiments, the reconstructed set of data of output block {circumflex over (x)} 202 includes a signal displayed as an image on monitor 47 of FIG. 1. In some embodiments, the reconstructed set of data of output block {circumflex over (x)} 202 includes a signal played as a sound segment on speakers 57 of FIG. 1.

[0178] In some embodiments, as shown in FIG. 9, the present invention is implemented as an apparatus 900, and for such embodiments, FIG. 9 represents a schematic hardware block diagram. For such embodiments, overcomplete expansion transformation circuit 910 converts input set of data x 201 into overcomplete expansion data y 917. Receiver circuit 920 generates received data y′ 927. In some embodiments, overcomplete expansion data y 917 and received data y′ 927 are streams of data that are moved or transmitted as they are created; in other embodiments, one or both are stored in a data storage circuit such as a computer memory; in yet other embodiments, overcomplete expansion data y 917 is stored on a storage medium such as a DVD, and receiver 920 reads from the medium. Iterative transformation circuit 990 converts received data y′ 927 into reconstructed data {circumflex over (x)} 202. In these embodiments, blocks 910 and 990 are implemented with suitable multiplier, adder, accumulator, storage and control circuits as are well know in the computer art. In some such embodiments, reconstructed data {circumflex over (x)} 202 is image data that is displayed on monitor 47.

[0179] In some embodiments, as shown in FIG. 10, the present invention is implemented as a method 1000, and for such embodiments, FIG. 10 represents a method flowchart diagram. For such embodiments, overcomplete expansion transformation module 1010 converts input set of data x 201 into overcomplete expansion data y 217. Erasure channel 222 passes some or all of data y to received data y′ 227. In some embodiments, overcomplete expansion data y 217 and received data y′ 227 are streams of data that are moved or transmitted as they are created; in other embodiments, one or both are stored in a data storage circuit such as a computer memory; in yet other embodiments, overcomplete expansion data y 217 is stored on a storage medium such as a DVD, and data y′ 227 is read from the medium. Iterative transformation module 1075, convergence tester module 1070, and reconstruction transformation module 1080 convert received data y′ 227 into reconstructed data {circumflex over (x)} 202. In these embodiments, blocks 1075, 1070, and 1080 are implemented with suitable software modules as are well know in the computer art to implement the present invention as described above. In some such embodiments, reconstructed data {circumflex over (x)} 202 is image data that is displayed by module 1085 onto a display monitor.

[0180] In some embodiments, there are two improvements that are, either together or individually, made to this basic algorithm. The first improvement that is made in some embodiments is to reconstruct to a point in the interior of the quantization cell, rather than to a corner. Hopefully, such a reconstruction will be closer to the cell centroid. For this purpose, modify the algorithm to gradually shrink P₁, . . . , P_Mtowards their centers (or approximate centroids) until there is no point of intersection. That is, run the basic algorithm to convergence, increase the lower limits, decrease the upper limits, and run the basic algorithm again to convergence. The limiting points for each run, say p₁(∞), . . . , P_M(∞), will be equal until there is no point of intersection between P₁, . . . , P_M, after which they will begin to diverge. Reconstruct {circumflex over (x)} from the last limiting point p₁(∞) for which p₁(∞) approximately equals p₂(∞), . . . , p_M(∞).

[0181] The second improvement that is made in some embodiments, when the number of received descriptions N′=|R₁|+ . . . +|R_M| is less than K, is to reconstruct missing components to their conditional expected values given the received descriptions. The basic algorithm already does this if y₁has a spherical density (e.g., if the components of y₁are independent identically distributed Gaussian random variables). The reason is that if y₁has a spherical density, then the subvector y₁″ consisting of the erased components {y_1,i}_i∉R1has a spherical conditional density given the received components {y_1,i}_i∉R1. Furthermore, given the received components {y_2,i}_i∉R2through {Y_m,i}_i∉RM, y₁″ must lie in some linear (|R₂|+ . . . +|R_M|) dimensional variety. Thus, the conditional density of y₁′ given all the received components is a spherical distribution in some linear variety with its mean at the point where the all-zero vector y₁′=0 projects onto the linear variety. Therefore, setting the missing components in the initial point p₁(0) in Step 1 of the basic algorithm will result in their being replaced, after projection in Step 3, by their conditional means given the received descriptions. Although in most circumstances the density of y₁is not spherical, it will be approximately spherical if F₁is a decorrelating transform and x is preconditioned such that it has zero mean and the variance of each y₁is constant. More precisely, if σ_i²is the variance of (F₁x)_i, then replace x by F₁⁻¹diag (σ₁⁻¹, . . . σ_K⁻¹) F₁(x−EX). The resulting vector y₁will have approximately spherical density.

[0182] Experimental Results

[0183] Results are presented using the given algorithm to reconstruct a K dimensional vector x. In one test, vector x is originally formed by taking the inverse DCT (discrete cosine transform; see the book by M. Vetterli and J. Kova{haeck over (c)}ević, Wavelets and Subband Coding, Prentice Hall, 1995; page 340, page 374; herein “Vetterli et al.” which provides a general reference for various types of transforms used in various embodiments of the present invention.) of a vector of transform coefficients y₁, which are in turn sampled from a Gaussian distribution with mean 0 and diagonal covariance with entries inversely proportional to frequency, i.e., Y_1,k˜N(0, σ²/(1+ck)), k=0, 1, . . . , k−1, where σ²and c are constants. The vector x is then transformed using a DCT for F₁and the identity transform for F₂, yielding y₁=DCT(x) and y₂=x. The transform coefficient vectors y₁and y₂are uniformly scalar quantized using a step size of Δ. In one experiment, K=512, σ²=1.0, Δ=0.1, and c=0.013, so that the variance of the last coefficient is σ²/8. Vetterli et al. (page 374) provides the following definitions of the DCT used in one embodiment of the present invention:

\begin{matrix} y_{0} = \frac{1}{\sqrt{(N)}} \sum_{n = 0}^{N - 1} x_{n} & (Equation 10) \\ y_{k} = \sqrt{\frac{2}{N}} \sum_{n = 0}^{N - 1} x_{n} \cos (\frac{2 π (2 n + 1) k}{4 N}), k = 1, \dots, N - 1 & (Equation 11) \end{matrix}

[0184] All of the quantized coefficients ŷ₁are transmitted, along with ¼ of the quantized coefficients ŷ₂(randomly selected), for a redundancy of {fraction (N/K)}=1.25. Of these 640 transmitted coefficients, random subsets are received. The reconstruction algorithm is run for each subset received. The performance of the algorithm is measured by averaging the reconstruction error over all subsets having the same number of coefficients. The performance of a comparable forward-error-correction system is also obtained. If the number of coefficients received in the FEC system is not enough to reconstruct the vector, then only the systematic portion of the received coefficients is used.

[0185] A plot of N″ vs. SNR 191 (signal-to-noise ratio) in dB (decibels) is shown in FIG. 4. There is a slight gain with the present system (as indicated by graph line 192) over the plain FEC system (as indicated by graph line 193) if all the coefficients are received. If fewer than K coefficients are received, then the present system always outperforms the FEC system. However, there is substantial loss if exactly K coefficients are received. This is due to the fact that the partition induced by the received coefficients is usually not cubic since there is a mixture of coefficients from the two transfonns. Also, the point found by the POCS algorithm is not necessarily the centroid of the cell. The exact value of the loss is somewhat arbitrary; performance of the FEC system when more than K coefficients are received can be made as high as desired by reducing Δ, whereas performance of even the optimal reconstruction when fewer than K coefficients are received reaches an upper bound independent of Δ.

[0186] FIG. 6 shows an image 593 of “Lena” with POCS. (Reconstruction from an overcomplete representation with redundancy 1.25, with ⅛ of the transmitted components missing.)

[0187] Experimental results are also provided for an image called Lena, using for F₁a three-level dyadic separable 2D (two-dimensional) wavelet transform based on the 9/7 filter of (M. Antonini, M. Barlaud, P. Mathieu, and I. Daubechies. Image coding using wavelet transform. IEEE Trans. Image Processing,1:205-221, April 1992; herein “Antonini et al.”), and using for F₂the separable 2D DCT. The coefficients are uniformly scalar quantized to a stepsize of 16. All of the coefficients of ŷ₁are transmitted, along with the coefficients of ŷ₂corresponding to the lower half horizontal frequencies and the lower half vertical frequencies, for a redundancy of N/K=1.25. Of these transmitted coefficients, ⅛ are erased at random. FIG. 6 shows the reconstruction. The reconstruction error in this example is 30.69 dB, compared to 15.68 dB if the same ⅛ of the wavelet coefficients are erased without benefit of being reconstructed from the extra coefficients from F₂, as shown in FIG. 5 (an image 592 of “Lena” having missing coefficients for ⅛ of the pixels). However, this compares to 37.77 dB if the wavelet coefficients y₁=F₁x are all received in their entirety.

[0188] In summary, the present invention provides a POCS-based algorithm for consistent reconstruction from multiple descriptions of overcomplete expansions. The algorithm operates in the data space RK rather than in the expanded space R^N, N>K. By constructing the frame from two complete transform bases, all projections can be expressed in terms of forward or inverse transforms. Since such transforms are usually efficient to compute, the present invention can perform the reconstruction much faster than with previous methods. Indeed, the method of one embodiment enables overcomplete frame expansions of an entire image, which has heretofore not been possible.

[0189] The present invention represents a POCS-based algorithm for consistent reconstruction of a signal xεR^Kfrom any subset of quantized coefficients yεR^Nin an N×K overcomplete frame expansion y=Fx, N=MK (where M is two or larger). By choosing the frame operator F to be the concatenation of two or more K×K invertible transforms, the projections may be computed in R^Kusing only the transforms and their inverses, rather than in the larger space R^Nusing the pseudo-inverse as proposed in earlier work. This enables practical reconstructions from overcomplete frame expansions based on wavelet, subband, or lapped transforms (one or more of which are used in some of the various embodiments of the present invention) of an entire image, which has heretofore not been possible.

[0190] Unequal Error Protection

[0191] The present invention also provides a method and apparatus for unequal error protection for multiple description source coding. By analogy, this can be compared with unequal error protection for error-correction codes. In particular, given a bit-plane representation of the coefficients in an overcomplete representation of the source, one embodiment of the present invention provides coding the most significant bits with the highest redundancy and the least significant bits with the lowest redundancy. In one embodiment, this is accomplished by varying the quantization stepsize for the different coefficients. Then, in some embodiments, the available received quantized coefficients are reconstructed using a method based on alternating projections onto convex sets (or other suitable method) as described above.

[0192] In multiple-description quantization using overcomplete (frame) expansions as described above, an input signal xεR^Kis represented by a vectory y=FxεR^N, where N>K. F is a so-called “frame operator,” whose N rows span R^K. The coefficients of y are scalar quantized to obtain ŷ and then are independently entropy coded and transmitted in (up to) N descriptions. The decoder receives descriptions of only N′≦N coefficients after potential erasures, and reconstructs the signal from the received descriptions. Each received description is an encoding of the fact that some coefficient y_ilies in a particular quantization interval, say [l_i,u_i).

[0193] In the description above, a practical method and system are presented for consistent reconstruction of x given any subset of the transmitted quantization intervals [l_i,u_i), where l≦Fx<u (pointwise). In some embodiments of the method and/or system, F has the form

F = [\begin{matrix} F_{1} \\ F_{2} \end{matrix}]

[0194] where F₁is a K×K transform, and F₂is another K×K transform, orthogonally related to F₁(i.e., F₂=QF₁for some orthogonal transform Q). For example, F₁may be a wavelet transform (over an image xεR^Ksuitably extended), and F₂may be the identity transform, another wavelet transform, or the same wavelet transform over a 1-pixel shift of the image. The algorithm used by the present invention reconstructs x from the received quantization intervals [l_i, u_i) using alternating projections onto convex sets. The notation “[l_i, u_i)” indicates the interval greater than or equal to l_iand less than u_i.

[0195] In the description above, multiple-description quantization is compared using the above reconstruction algorithm to a systematic error-correction code, in a hypothetical transmission system. In the hypothetical transmission system, the coefficients in the K-dimensional vectory, y₁=F₁x are uniformly scalar quantized to B bits of precision, and are separately transmitted in K B-bit descriptions. Redundancy is provided by transmitting an additional N−K B-bit descriptions. In the multiple-description case, the additional descriptions encode the first N−K coefficients of the real vector y₂=F₂x uniformly scalar quantized to B bits of precision. In one embodiment, reconstruction is performed by the aforementioned algorithm. In the error-correction case, the additional packets contain the N−K B-bit parity check symbols from the finite field GF(2^B) generated by a systematic (N, K) error-correction code. Reconstruction is given by {circumflex over (x)}=F₁⁻¹ŷ₁when N−K or fewer packets are lost. When more than N−K packets are lost, reconstruction is given by the same formula, with the lost coefficients set to zero in ŷ₁. Results are generated in both cases by simulating random reception of N′ out of N descriptions (in repeated trials) for N′=0, . . . , N. FIGS. 11 and 12 show PSNR (peak signal-to-noise ratio) results for Lena comparing multiple-description and error control coding, with FIG. 11 showing reconstruction results for equal protection and FIG. 12 showing reconstruction results for unequal error protection. FIG. 11 shows the expected squared reconstruction error in each case as a function of N′, for the image called “Lena”. Here, F₁is the 9/7 biorthogonal wavelet transform and F₂is the separable full-image DCT. When N′<K, MD (Multiple Description) quantization outperforms error-correction coding (ECC) by up to 14 dB, while for N≧K (except at N′=N), error-correction coding outperforms MD quantization by up to 3 dB. The 3 dB loss for MD quantization is apparently due to the fact that the quantization cells in the signal domain are not cubic since the received coefficients correspond to a mixture of basis functions from the transforms F₁and F₂and hence are not orthonormal.

[0196] The above comparison is somewhat unfair to the error-correction coding approach, because, as is well known, better performance can be achieved by protecting different bits with different redundancies. If the bits of each quantized coefficient are labeled 1, . . . , B from the least to the most significant, then loss of the bth bit increases the squared error of the reconstruction in proportion to 2^2b. Indeed, if ε_K(N) is the probability that an (N, K) code cannot recover a symbol, then the optimal redundancy N/K for coding the bth bit is approximately given by 2^2bε_K(N)=constant, where ε′_K(N) is the slope of ε_K(N) at N. This is a classic bit-allocation problem whose solution gives the optimal allocation of parity bits to different bit planes of the transform coefficients so as to minimize the expected squared-reconstruction error. Unequal error protection has been employed in other contexts.

[0197] This section describes embodiments that include unequal error protection for multiple-description quantization. One embodiment of the present invention is illustrated in FIG. 13 that shows components of the vectors y₁=F₁x and y₂=F₂x depicting which bits are to be transmitted in one embodiment of unequal error protection. The components of the vectors y₁=F₁x and y₂=F₂x are depicted in a sequence from left to right, while the bits of each component are depicted from most to least significant, top to bottom. The shaded region shows which bits are to be transmitted. All of the bits from the coefficients y₁are transmitted, while only some of the bits from the coefficient y₂are transmitted. Looked at in one way, each bit plane (or row of bits) has its own redundancy. The most significant bits have the highest redundancy while the least significant bits have the lowest redundancy. Looked at in another way, each coefficient is quantized to its own number of bits. All of the coefficients of y₁are quantized to B bits, but the ith coefficient of y₂is quantized to b_i≧B bits, with b_idecreasing to 0 as the index i of the coefficient increases. The quantized coefficients are transmitted as usual; a random number of them are erased. In one embodiment, the reconstruction procedure, using the iterative algorithm of the present invention, does not change, except for the fact that the received quantization intervals [l_i, u_i) now vary in width to reflect the unequal stepsizes for the different coefficients. These stepsizes need not be related by powers of two. Thus, in some embodiments, the stepsizes are related by powers of two; and in other embodiments, the stepsizes are not related by powers of two.

[0198] FIG. 12 shows the expected squared reconstruction error of multiple-description quantization and error control coding analogous to FIG. 11, except that the redundant bits have been reallocated to different bit planes to provide unequal error protection. The gains are about 1 to 2 dB over the equal error protection case. The allocation of redundant bits is the same in both MD and ECC cases. Reconstruction in the error control case assumes that if a bit is lost, then all less significant bits within the same coefficient are lost as well.

[0199] In some embodiments of the present invention, algorithms for determining the optimal allocation of redundant bits in the multiple-description case are determined empirically.

[0200] It should be noted that unequal error protection in multiple description coding by varying the quantization precision across coefficients can be used very generically, with any kind of overcomplete transform F and any kind of reconstruction method. Two instances of prior art in which such unequal error protection has been used are the following.

[0201] In W. Jiang and A. Ortega, “Multiple description coding via polyphase transform and selective quantization,” SPIE Conf. on Visual Communications and Image Processing, January 1999, F is the concatenation of two identical K×K transforms, F₁and F₂. The coefficients in F₂are quantized to lower resolution than the coefficients in F₁. Reconstruction is performed simply by replacing the coefficients missing from F₁with the corresponding low-resolution coefficients from F₂.

[0202] In A. C. Miguel, A. E. Mohr, and E. A. Riskin, “SPIHT for generalized multiple description coding,” submitted to IEEE Int'l Conf. on Image Processing, Kobe, Japan, 1999, F is the concatenation of four identical K×K transforms, F₁, . . . , F₄. The coefficients in F₁are quantized to full precision (10 bit planes). The coefficients in F₂, F₃, and F₄, are quantized to 7, 5, and 4 bit planes, respectively, and grouped so that corresponding coefficients do not share the same binary description. Reconstruction is performed simply by replacing the coefficients missing from F₁with the corresponding low resolution coefficients from F₂, F_3,or F₄.

[0203] One feature that distinguishes this invention from prior art is that in this invention the transforms F_i, for i>1, do not equal F₁. In other words, if F is the overcomplete frame expansion for the multiple description system of this invention, then the last N−K coefficients are not equal to the first K coefficients, before quantization.

CONCLUSION—UNEQUAL ERROR PROTECTION

[0204] One aspect of the present invention provides a computerized method of communicating a set of data x having K dimensions. The method enhances the recoverability of a set of data x having K dimensions for transmission on an erasure channel. The method includes generating an expansion y of the set of data x using an N×K multiple description overcomplete expansion y=Fx, wherein N>K and wherein at least some of the last N−K rows of F are not duplicative of any of the first K rows. The method also includes quantizing the coefficients ofy to unequal resolution.

[0205] In some embodiments, the unequal quantization resolution is based on a significance of data to its recovery.

[0206] In some embodiments, the set of data has a bit-significant representation, and the overcomplete expansion of the set of data provides coding most significant bits with a higher redundancy and least significant bits with a lower redundancy.

[0207] In some embodiments, the set of data is a source image having a bit-plane representation, and wherein coefficients of the overcomplete expansion of the source image provide coding the most significant bits with a higher redundancy and the least significant bits with a lower redundancy. In some such embodiments, the method further includes varying the quantization stepsize for the different coefficients. In some such embodiments, the method further includes receiving at least some of the quantized coefficients and decoding the received quantized coefficients using a method based on alternating projections onto convex sets in order to generate a recovered set of data.

[0208] One embodiment of the present invention provides a computer-readable medium having instructions stored thereon for causing a computer to perform one or more of the methods just described.

[0209] Another aspect of the present invention provides a computerized system for enhancing the recoverability of a set of data for transmission on an erasure channel. The system includes means for generating an overcomplete expansion of the set of data wherein unequal encoding is employed. In one embodiment, the unequal encoding is based on a significance of data to its recovery. In another embodiment, the set of data has a bit-significant representation, and the overcomplete expansion of the set of data provides coding most significant bits with a higher redundancy and least significant bits with a lower redundancy.

[0210] In some embodiments, the set of data is a source image having a bit-plane representation, and wherein the coefficients of the overcomplete expansion of the source image provide coding the most significant bits with a higher redundancy and the least significant bits with a lower redundancy.

[0211] Some embodiments of the system further include means for varying the quantization stepsize for the different coefficients. Other embodiments further include means for receiving at least some of the quantized coefficients and means for decoding the received quantized coefficients using a method based on alternating projections onto convex sets. Still other embodiments further include a computer-readable medium having instructions stored thereon for causing a computer to operatively control the above mentioned means for generating the overcomplete expansion.

[0212] Another aspect of the present invention provides a computerized method of reconstructing data after a transmission on an erasure channel. This method includes providing an initial set of data x having K dimensions, overcompletely transforming the set of data x to an overcomplete expansion set of data y having N dimensions, where N>K, and quantizing the coefficients ofy to unequal resolution.

[0213] In some embodiments the set of data is a source image having a bit-plane representation, and the quantized coefficients of the overcomplete expansion of the source image provide coding the most significant bits with a higher redundancy and the least significant bits with a lower redundancy.

[0214] Some embodiments of the method further include varying the quantization stepsize for the different coefficients.

[0215] Some of these method embodiments further include transmitting the set of data y on the channel, receiving a set of data y′ from the channel, and iteratively transforming the received set of data y′ using projections onto convex sets that use at least two different K×K transforms including:

[0216] (a) initializing t;

[0217] (b) setting an initial point in F₁P;

[0218] (c) transforming p₁(t) into the coordinate system of F₂;

[0219] (d) projecting p₂(t) onto F₂Q;

[0220] (e) transforming q₂(t+1) into the coordinate system of F₁;

[0221] (f) projecting q₁(t+1) onto F₁P;

[0222] (g) checking for convergence and if not sufficiently converged then iterating (c) through (g); and

[0223] (h) reconstructing {circumflex over (x)}=F_n⁻¹p_n(t+1) in order to generate a recovered set of data; wherein:

[0224] F₁and F₂are each K×K transforms, and n=1 or 2.

[0225] In some of these method embodiments

[0226] (a) initializing further includes initializing t=0;

[0227] (b) starting further includes setting an initial point p₁(0) such that for each index i,

p_{1, i} (0) = {\hat{y}}_{1, i} if i \in R_{1} and

p_{1, i} (0) = 0 if i \notin R_{1};

[0228] (c) transforming P₁(t) further includes transforming p₁(t) into the coordinate system of F₂using p₂(t)=F₂(F₁⁻¹p₁(t))=F₁₂p₁(t);

[0229] (d) projecting p₂(t) further includes projecting p₂(t) onto F₂Q such that for each index i:

q_{2, i} (t + 1) = \min {\max {p_{2, i} (t), l_{2, i}}, u_{2, i}} if i \in R_{2} and

q_{2, i} (t + 1) = p_{2, i} (t) if i \notin R_{2};

[0230] (e) transforming q₂(t+1) further includes transforming q₂(t+1) into the coordinate system of F_{using q}₁(t+1)=F₁(F₂⁻¹q₂(t+1))=F₂₁q₂(t+1);

[0231] (f) projecting q₁(t+1) further includes projecting q₁(t+1) onto F₁P such that for each index i:

p_{1, i} (t + 1) = \min {\max {q_{1, i} (t + 1), l_{1, i}}, u_{1, i}} if i \in R_{1} and

p_{1, i} (t + 1) = p_{2, i} (t) if i \notin R_{1}; and

[0232] (g) checking further includes checking for convergence by testing ∥p_n(t+1)−p_n(t)∥²>ε, and if so then setting t←t+1 and iterating (c) through (g); and

[0233] (h) reconstructing {circumflex over (x)}=F_n⁻¹p_n(t+1) wherein n=1 or 2, and wherein:

[0234] F₁and F₂are each K×K transforms,

[0235] F₁⁻¹is the inverse transform of F₁, F₂⁻¹is the inverse transform of F₂, p1 ŷ₁is a quantized version of y₁,ŷ₂is a quantized version of y₂, such that ŷ componentwise lies between respective lower and upper quantization cell boundary vectors l₁≦ŷ₁≦u₁and ŷ₂componentwise lies between respective lower and upper quantization cell boundary vectors l₂≦ŷ₂≦u₂,

[0236] R₁⊂{1, . . . ,K}is the set of indices of the descriptions received of ŷ₁, and

[0237] R₂⊂{1, . . . ,K} is the set of indices of the descriptions received of ŷ₂.

[0238] In some of the method embodiments, the methods are performed in the order listed. Other embodiments use different orders.

[0239] This application is intended to cover any adaptations or variations of the present invention. It is manifestly intended that this invention be limited only by the claims and equivalents thereof.

Claims

Certificate of Correction present

The claims shown below may not include correction changes. Use the corrected PDF for authoritative text.

What is claimed is:

1. A computerized method of enhancing the recoverability of a set of data x having K dimensions for transmission on an erasure channel, the method comprising:

generating an overcomplete expansion y of the set cf data x using an N×K multiple description overcomplete expansion y=Fx, wherein N>K and wherein at least some of the last N−K rows of F are not duplicative of any of the first K rows; and

quantizing the coefficients of y to unequal resolution.

2. The method of claim 1, wherein the unequal quantization resolution is based on a significance of coefficients to data recovery.

3. The method of claim 1, wherein the set of data has a bit-significant representation, and wherein the overcomplete expansion of the set of data provides coding most significant bits with a higher redundancy and least significant bits with a lower redundancy.

4. The method of claim 1, wherein the set of data is a source image having a bit-plane representation, and wherein coefficients of the overcomplete expansion of the source image provide coding the most significant bits with a higher redundancy and the least significant bits with a lower redundancy.

5. The method of claim 4, further comprising:

varying the quantization stepsize for the different coefficients.

6. The method of claim 5, further comprising:

receiving at least some of the quantized coefficients; and

decoding the received quantized coefficients using a method based on alternating projections onto convex sets in order to generate a recovered set of data.

7. A computer-readable medium having instructions stored thereon for causing a computer to perform the method of claim 1.

8. A computerized system for enhancing the recoverability of a set of data for transmission on an erasure channel, the system comprising:

a component for generating a multiple-description overcomplete expansion of the set of data wherein unequal encoding is employed.

9. The system of claim 8, wherein the unequal encoding is based on a significance of data to its recovery.

10. The system of claim 8, wherein the set of data has a bit-significant representation, and wherein the overcomplete expansion of the set of data provides coding most significant bits with a higher redundancy and least significant bits with a lower redundancy.

11. The system of claim 8, wherein the set of data is a source image having a bit-plane representation, and wherein the coefficients of the overcomplete expansion of the source image provide coding the most significant bits with a higher redundancy and the least significant bits with a lower redundancy.

12. The system of claim 11, further comprising

means for varying the quantization stepsize for the different coefficients.

13. The system of claim 12, further comprising

means for receiving at least some of the quantized coefficients; and

means for decoding the received quantized coefficients using a method based on alternating projections onto convex sets.

14. The system of claim 8, further comprising a computer-readable medium having instructions stored thereon for causing a computer to operatively control the means for generating the overcomplete expansion.

15. A computerized method of reconstructing data after a transmission on an erasure channel, the method comprising:

providing an initial set of data x having K dimensions;

overcompletely transforming the set of data x to a multiple-description overcomplete expansion set of data y having N dimensions, where N>K; and

quantizing the coefficients of y to unequal resolution.

16. The method of claim 15, wherein the set of data is a source image having a bit-plane representation, and wherein coefficients of the overcomplete expansion of the source image provide coding the most significant bits with a higher redundancy and the least significant bits with a lower redundancy.

17. The method of claim 16, further comprising:

varying the quantization stepsize for the different coefficients.

18. The method of claim 17, further comprising:

transmitting the set of data y on the channel;

receiving a set of data y′ from the channel; and

iteratively transforming the received set of data y′ using projections onto convex sets that use at least two different K×K transforms including:

(a) initializing t;

(b) setting an initial point in F₁P;

(d) projecting p₂(t) onto F₂Q;

(e) transforming q₂(t+1) into the coordinate system of F₁;

(f) projecting q₁(t+1) onto F₁P;

(g) checking for convergence and if not sufficiently converged then iterating (c) through (g); and

(h) reconstructing {circumflex over (x)}=F_n⁻¹p_n(t+1) in order to generate a recovered set of data; wherein:

F₁and F₂are each K×K transforms, and n=1 or 2.

19. The method of claim 15, further comprising:

transmitting the set of data y on the channel;

receiving a set of data y′ from the channel; and

iteratively transforming the received set of data y′ using projections onto convex sets that use at least two different K×K transforms including:

(a) initializing t;

(b) setting an initial point in F₁P;

(d) projecting p₂(t) onto F₂Q;

(e) transforming q₂(t+1) into the coordinate system of F₁;

(f) projecting q₁(t+1) onto F₁P;

(g) checking for convergence and if not sufficiently converged then iterating (c) through (g); and

(h) reconstructing {circumflex over (x)}=F_n⁻¹p_n(t+1) in order to generate a recovered set of data; wherein:

F₁and F₂are each K×K transforms, and n=1 or 2.

20. The method of claim 19, wherein:

and wherein:

(a) initializing further includes initializing t=0;

(b) starting further includes setting an initial point p₁(0) such that for each index i,

p_{1, i} (0) = {\hat{y}}_{1, i} if i \in R_{1} and

p_{1, i} (0) = 0 if i \notin R_{1};

(c) transforming p₁(t) further includes transforming p₁(t) into the coordinate system of F₂using p₂(t)=F₂(F₁⁻¹p₁(t))=F₁₂p₁(t);

(d) projecting p₂(t) further includes projecting p₂(t) onto F₂Q such that for each index i:

q_{2, i} (t + 1) = \min {\max {p_{2, i} (t), l_{2, i}}, u_{2, i}} if i \in R_{2} and

q_{2, i} (t + 1) = p_{2, i} (t) if i \notin R_{2};

(e) transforming q₂(t+1) further includes transforming q₂(t+1) into the coordinate system of F₁using q₁(t+1)=F₁(F₂⁻¹q₂(t+1))=F₂₁q₂(t+1);

(f) projecting q₁(t+1) further includes projecting q₁(t+1) onto F₁P such that for each index i:

p_{1, i} (t + 1) = \min {\max {q_{1, i} (t + 1), l_{1, i}}, u_{1, i}} if i \in R_{1} and

p_{1, i} (t + 1) = p_{2, i} (t) if i \notin R_{1}; and

(g) checking further includes checking for convergence by testing ∥p_n(t+1)−p_n(t)∥²>ε, and if so then setting t←t+1 and iterating (c) through (g); and

(h) reconstructing {circumflex over (x)}=F_n⁻¹p_n(t+1), wherein n=1 or 2,

F₁and F₂are each K×K transforms,

F₁⁻¹is the inverse transform of F₁, F₂⁻¹is the inverse transform of F₂,

ŷ₁is a quantized version of y₁, ŷ₂is a quantized version of y₂, such that ŷ₁componentwise lies between respective lower and upper quantization cell boundary vectors l₁≦ŷ₁≦u₁and ŷ₂componentwise lies between respective lower and upper quantization cell boundary vectors l₂≦ŷ₂≦u₂,

R₁⊂{1, . . . ,K} is the set of indices of the descriptions received of ŷ₁, and

R₂⊂{1, . . . ,K} is the set of indices of the descriptions received of ŷ₂.

21. The method of claim 20, wherein the (a) initializing t, (b) setting an initial point in F₁P, (c) transforming p₁(t) into the coordinate system of F₂, (d) projecting p₂(t) onto F₂Q, (e) transforming q₂(t+1) into the coordinate system of F₁, (f) projecting q₁(t+1) onto F₁P, (g) checking for convergence and if not sufficiently converged then iterating (c) through (g); and (h) reconstructing {circumflex over (x)}=F_n⁻¹p_n(t+1) are performed in the order listed.

22. The method of claim 19, wherein: