Data compression method and system using globally optimal scalar quantization

Abstract

A signal encoding/decoding system and method are provided, where a signal quantizer is designed by minimizing a target function in accordance with a rate/distortion tradeoff. The system and method can be applied to give lossy compression of data for graphic or music files and also to smart-server applications, in accordance with the available bandwidth.

Description

BACKGROUND OF THE INVENTION

1. Field of the Invention

The present invention relates to a data compression method and system. More in particular, it relates to a method for designing an optimal scalar quantizer to be used in communications or storage systems.

In a communication system, the information to be sent along a channel is usually encoded by means of a source coder, which compresses the data to be sent. The source coder employs a source code that can either be lossless or lossy. A lossless source code is completely reversible, while a lossy source code creates a distortion of the signal to be transmitted. Any lossy source code can be described as a quantizer followed by a lossless compression algorithm. The quantizer performs an approximation step and the lossless compression algorithm performs a lossless compression step.

Quantizers that operate on blocks of source output are called vector quantizers. Quantizers that quantize each output separately are called scalar quantizers. In scalar quantization, each single-source output is quantized to one of some number of levels, and these levels are encoded into a binary sequence. See, for example, “Communication Systems Engineering,” John G. Proakis and Masoud Salehi, Prentice-Hall Inc., 1994, p. 246.

For example, the initial source alphabet can be a list of 256 brightness levels numbered from 0 to 255 in a gray scale image. Each symbol can have a certain relative frequency, or probability, so that a list of symbol/probability pairs can be defined, forming a discrete probability function, also known as probability mass function or pmf. A scalar quantizer would approximate the initial source alphabet with a smaller set, for example, the set {10, 35, 49, 102} by mapping each number between 0 and 255 to either 10, 35, 49 or 102. A new probability function would arise, given by the relative frequencies of 10, 35, 49 and 102 in the approximation. The “distortion measure” quantities indicate how badly the data is distorted by the chosen mapping.

The output of the quantizer is assumed to go to a lossless compression algorithm, which introduces no further approximations, but produces a more compact representation of the data by taking advantage of its redundancy. The quantizer mapping is chosen with the aim of both keeping the distortion low and allowing a compact representation. The “rate” describes how compact the representation is, with lower rate indicating a shorter data representation. The shorter data descriptions require less memory to store and less time to transmit. Generally, coarser approximations yield higher distortion and lower rate.

There are two important rate measures: in “fixed rate” scenarios, the number of symbols for the approximation is fixed, and each approximation has the same description length. Therefore, the rate does not enter into consideration, and all that remains in quantizer design is the choice of approximation symbols that yield the least distortion, i.e. the least approximation error. In “variable rate” scenarios, the number of approximating symbols is not fixed beforehand and each approximation can have a different description length.

In some systems, an optional initial transform step is provided, where an initial reversible transformation is performed as preprocessing. The code then quantizes the transformed data. The reconstruction process replaces each description by its corresponding reconstruction and then reverses the transformation.

When only one approximation of the original data is given, the quantizer is called a single-description quantizer. However, there is a need to design quantizers not limited to single-description scenarios, but also capable of handling other coding scenarios, like multiple description and multiple resolution coding.

Multiple description refers to the simultaneous production of two or more approximations of the original data. Some approximations are given independent descriptions, while others are reached by combining two or more of the independent descriptions. Reference can be made, for example, to a communications system where more than one path is available, like the case, for example, of one transmitter and multiple receiving antennas. In such case, the different descriptions will travel independently, and some of them may not be received at all.

Multiple description (MD) can be general or restricted. In general-MD, all scenarios are a-priori possible. In restricted-MD, some combinations are known to be impossible. Therefore, in restricted-MD, the sender can make certain assumptions that reduce the number of scenarios to be considered and simplify the code. A special case of restricted-MD is called multi-resolution. In multi-resolution, each description is received only if the previous ones have been received, so that description 3 cannot arrive without description 1 and description 2. A specialized version of the method according to the present invention runs much faster in the multi-resolution case.

Throughout the description of the present invention, reference will be made to the enclosed Annex A, which makes part of the present disclosure. Annex A also contains a reference to twenty prior art publications. The same notation [1] . . . [20] used for those references in Annex A will be used throughout the description.

2. Description of the Prior Art

Prior art approaches in globally optimal scalar quantizer design for fixed-rate, single-encoder, single-decoder codes for finite-alphabet sources are known. See, for example, references [1], [2], [3], and [4] of Annex A.

A quantizer design algorithm for variable-rate scalar quantization is also known from reference [2] of Annex A, but is very slow.

Additionally, algorithms are known which handle both fixed- and variable-rate, single or multiple descriptions, by iterative methods, where the initial step is that of providing an arbitrary solution and then obtaining a better solution after each iteration. However, these algorithms do not necessarily yield an optimal quantizer, and have no bound on the running time.

Therefore, there is a need for a method for designing an optimal quantizer for signal encoding/decoding which can be applied to the multiple environments described above. Such method must also have a polynomial complexity.

SUMMARY OF THE INVENTION

The present invention overcomes the above limitations and problems by providing a method which finds a quantizer in single-description and multiple-description scenarios with polynomial complexity. Under appropriate conditions, explained throughout the following description, the quantizer is the optimal quantizer. In particular, the quantizer according to the present invention is optimal when it provides the optimum rate/distortion tradeoff. An optimum rate/distortion combination is a combination which makes the rate no larger than some specified allowed value and reduces distortion as much as possible. In particular, the improvement in the quantizer distortion associated with a small increase in rate is matched to a desired target value. The target value is chosen so that the distortion is minimized.

According to a first aspect of the present invention, a signal encoding system having a coder for encoding signals is provided, the coder including a quantizer and a lossless coder, the quantizer producing a distortion on the signals to be encoded, wherein the quantizer is an optimal quantizer under a rate/distortion tradeoff, so that the distortion is minimized when the rate of the optimal quantizer is no larger than a specified value.

According to a second aspect of the present invention, a signal decoding system having a decoder for decoding signals is provided, the decoder including a lossless decoder and an inverse quantizer wherein the inverse quantizer is an optimal quantizer under a rate/distortion tradeoff, so that the distortion is minimized when the rate of the optimal quantizer is no larger than a specified value.

According to a third aspect of the present invention, a signal encoding method for use in a signal encoding system is provided, comprising the steps of: providing a source alphabet containing a set of source characters; approximating the set of source characters by designing a quantizer and applying the designed quantizer to the set of source characters, thus producing a distortion on the set of source characters; and applying a compression algorithm to the smaller set of characters, wherein the step of designing a quantizer comprises the steps of: defining a target function depending on the distortion on the set of source characters; and minimizing the target function.

According to a fourth aspect of the present invention, a client-server system comprising a server and a plurality of clients is provided, wherein when a client of the plurality of clients requests data to the server, the server provides the client with an approximation of the requested data, the approximation being based on a quantization process of the requested data, thus producing distortion on the requested data, the quantization process approximating the requested data to a smaller number of data and comprising a step of designing a quantizer by defining a target function and minimizing the target function.

The method according to the present invention can be applied to a variety of coding scenarios (single-description; general or restricted multiple description; and multi-resolution), under a variety of distortion measures (for example, mean-squared error; absolute error; other monotonic error measures) and under fixed- or variable-rate, in a bounded and reasonable amount of time (polynomial complexity).

According to a first step of the present invention, the quantizer design problem is transformed into a graph problem. In a further step, graph algorithms are used to solve the graph problem; and, in a third step, the graph solution is transformed into a quantizer, which represents the solution to the original problem.

More specifically, in the first step, a weighted directed acyclic graph (WDAG) is built, based on the parameters of the problem. A graph consists of a set of nodes, and a set of edges between nodes. The graph used in the preferred embodiment of the present invention is directed because each edge has an arrow (orientation) on it; weighted because each edge has a weight; and acyclic because there are no circuits.

In the second step, the method according to the present invention uses either the “single-source WDAG shortest path algorithm” or the “all-pairs WDAG shortest path algorithm” to obtain a path between two special nodes of the graph such that the sum of the weights of the edges that make up the path is minimal. Both kinds of the “shortest path algorithm” employ a topological sort algorithm, known as such in the prior art. Finally, the path is “decoded” to obtain a “partition” of the source alphabet. Such partition defines the quantizer.

The quantizer can then be combined with any lossless compression algorithm and optionally with a reversible transformation to yield a lossy compression scheme.

A main advantage of the method and system according to the present invention is that the time complexity of the algorithm it employs is polynomial. The time complexity of an algorithm is the amount of time (or number of steps) it takes to complete, as a function of the size of the input. The input size in the present invention is the number of symbol/probability pairs. The method according to the present invention runs in polynomial time in all embodiments. In the enclosed Annex A the degree of the polynomial is computed for each specific case. It should be noted that the polynomial time complexity of the method according to the present invention is a relevant result, because a brute-force method (examining all quantizers and choosing the best one) will take exponential time, because the number of possible quantizers for N symbols is 2 N .

A further advantage of the method is that it yields a code that is guaranteed to be the best of all possible codes with “contiguous codecells.” For fixed- and variable-rate single description, no other algorithm can yield better codes.

Since the method according to the present invention makes use of a trade-off parameter between rate and distortion, such method is particularly well suited for “smart server” applications where a client requests the data and the server delivers a suitable approximation, based on the available bandwidth, so that more rate and less distortion for large bandwidth, and less rate and more distortion for small bandwidth, are provided.

Examples of applications of the method according to the present invention could include a digital camera, a hand-held scanner, a medical image database or any other image server, where the server or device can store images in an efficient proprietary format internally to maximize usage of its limited memory, and produce a compressed file for download when required.

Additionally, a multiple description quantizer design is useful for “diversity coding” in wireless communications where more than one channel may be available and in packet networks like the Internet.

The quantizer designer method according to the present invention can be used as part of a lossy compression scheme like compression algorithms for images, video and audio. Additionally, the method according to the present invention can work with any kind of discrete-time signal, one-dimensional or multi-dimensional, like for example a sound signal, an image signal, or a time-series signal.

BRIEF DESCRIPTION OF THE DRAWINGS

The present invention will be understood and appreciated more fully from the following detailed description taken in conjunction with the drawings in which:

FIG. 1 shows a schematic representation of a digital communication system;

FIGS. 2 and 3 show flowcharts describing the generic operation of the method according to the present invention;

FIG. 4 shows a generic client-server environment, to which the present invention can be applied;

FIG. 5 shows a graph and a partition extension rule used in solving the 2DSQ method; and

FIG. 6 shows a schematical representation of the “decoder side-information” scenario.

Annex A, pages A1 to A17, is also enclosed to the present description and forms an integral part of the application as filed.

DETAILED DESCRIPTION OF THE INVENTION

Introduction

FIG. 1 shows a schematic representation of an encoding/decoding system, like for example a communications system or a storage system. A source coder 1 receives a signal from a signal source 2 and encodes the signal. Once encoded, the signal is optionally fed to a channel coder 3 which further encodes the signal and then input either to a transmission channel 4 or to a storage device, also represented by element 4 , like, for example, a memory of a digital imaging or music device. Once the encoded signal has been transmitted or stored, the signal is optionally initially decoded by a channel decoder 5 and then sent to a source decoder 6 , which performs an operation which is inverse to the operation of the source coder 1 , and finally sent to a receiver 7 . The source coder 1 comprises a quantizer 11 and a lossless coder 12 , whose input is connected to the output of the quantizer 11 . The source decoder 6 correspondingly comprises an inverse quantizer 61 and a lossless decoder 62 . Optionally, a reversible transform 13 and an inverse transform 63 are present.

FIGS. 2 and 3 show flowcharts describing the generic operation of the method according to the present invention.

In step S 1 of FIG. 2 , a source alphabet is provided, corresponding to the signal output by the signal source 2 of FIG. 1 . In step S 2 , a quantizer is designed. In step S 3 the designed quantizer is used, by approximating the set of source characters to a smaller set of characters. Finally, in step S 4 , a compression algorithm, corresponding to the lossless coder 12 of FIG. 1 , is applied to the quantizer data.

Steps S 21 -S 24 of FIG. 3 further define the step S 2 of FIG. 2 . In step S 21 , a target function is defined. In step S 22 , a graph, preferably a weighted acyclic graph, is built based on the data of the problem, and in step S 23 the target function is minimized by operation on the graph built in the previous step. In step S 24 , the solution to the graph problem is translated into a solution to the original problem (optimal quantization). Each path between the initial vertex and the final vertex of the partial RD graph corresponds to a partition of the source alphabet, or a set of side partitions for multiple decoder systems. The shortest path corresponds to the optimal partition. The optimal partition uniquely identifies the optimal quantizer.

In single description scenarios (called ECSQ), a path is a concatenation of arcs, and a partition is a concatenation of codegroups. By way of construction, an arc between vertices i and j corresponds to a codegroup comprising symbols i+1 . . . j (or grid cells i+1 . . . j if a “coarse grid” is used to a obtain a fast approximation). This corresponds directly to the optimal partition once the shortest path is known. For example, the Figure at page A4 of Annex A shows that the path 0→2→3 corresponds to the partition {{1,2}, {3}}.

For multiple description scenarios, each vertex contains a “tuple” (i.e. pair, triplet, 4-tuple, 5-tuple etc.) of labels. For example, in a 3DSQ scenario, each vertex contains three labels. For variable-rate general-3DSQ, these labels will be integers. For restricted-3DSQ or for fixed rate codes, each label might in turn be a pair or triplet of integers, and will be called “double-labeling.”

In the 3DSQ scenario, each of the three labels per vertex will contain one integer which is important—the “end-of-codegroup” counter. Additionally, each label may contain a “beginning-of-codegroup” counter for restricted-3DSQ, and a “number-of-codegroups” counter for fixed rate codes. To obtain the three side partitions corresponding to the three descriptions, a list of the “end-of-codegroup” counters of the vertices in the path is written down, for example:

0 0 0

1 0 0

1 2 0

1 2 1

3 2 1

3 3 1

3 3 3

By looking down the columns, three paths are obtained: 0→1→3 for the first column, which means that the first side partition (for the first partition) is {{1}, {2, 3}}, like in the ECSQ case. The second column reads 0→2→3, which means that the second side partition (for the second partition) is {{1, 2 }, {3}}. Finally, the third column gives 0→1→3 and the third side partition {{1}, {2, 3}}. It should be noted that in each step, only one of the three paths is extended (in terms of the original problem, only a single side codegroup is added to one of the side partitions, according to the “extension rule”). Thus column one which looks like 0 1 1 1 3 3 3 should be read as 0→1→3 as done previously.

FIG. 4 shows a generic client-server environment to which the present invention can be applied.

The following description will explain, in detail, steps S 21 to S 24 of FIG. 3 .

Problem Set-Up

In prior art fixed-rate coding, the number of codewords is fixed, all codewords are described at the same rate, and the design goal is to minimize the code's expected distortion with respect to a given probability mass function (pmf) or training set. See references [1], [2], [3], and [4] cited in Annex A.

The method according to the present invention will assume the distortion measure to be non-negative, and the expectation relating to the expected rates to be taken relative either to a known pmf on a finite alphabet or to a finite training set on an arbitrary alphabet. The alphabet will be written as 1 . . . N and the histogram describing the known pmf or the empirical estimate based on the training data will be denoted by a histogram p[1] . . . p[N].

Once either a quantizer encoder or a quantizer decoder is designed, optimization of the remaining portion is straightforward. The present invention will focus on the design of a quantizer encoder. The quantizer encoder according to the present invention will be able to be applied not only to single-resolution scalar quantizers, but also to multi-resolution and multiple description scalar quantizers. Although design of a quantizer decoder will not be described here in detail, such design will be straightforward for the person skilled in the art upon reading of the present description.

Designing an encoder on a finite alphabet is equivalent to designing one or more partitions on the source alphabet 1 . . . N, where each set in a partition describes a codegroup, i.e. all symbols given the same binary description. A traditional scalar quantizer encoder requires one partition. An M-description scalar quantizer encoder uses M partitions, one for each description. An M-resolution scalar quantizer encoder similarly requires M partitions.

The method according to the present invention finds, in each coding scenario, the optimal quantizer among all quantizers with contiguous codecells. The collection of allowed indices to which the space of possible data values is mapped by the quantizer is a set of reproduction values, called codewords. The set of codewords is called a codebook. The set of alphabet symbols that are mapped to a certain codeword is called the codecell associated with that codeword. When the source alphabet is a discrete subset of some continuous alphabet, the discrete codecells are called codegroups.

A quantizer with contiguous codecells will be called a regular quantizer. See the section “Codecell Contiguity” of Annex A, page A2. See also section VII of Annex A, pages A14-A17, where a more detailed discussion on the relationship between true optimality and regular or non-regular quantizers.

The discrete probability function is represented by histogram p[1] . . . p[N]. Given the contiguity assumption, each codegroup on the source alphabet corresponds to a segment on the probability signal. Since the performance of a partition (measured by the expected rate and distortion of the corresponding quantizer encoder and its optimal decoder) is a function of the probability signal, the partition-design problem on alphabet 1 . . . N can be treated as a segmentation problem on signal p[1] . . . p[N], as also indicated in the section “Segmentation and Quantization” of Annex A, page A2. In particular, the optimal quantization problem can be cast as an instance of optimal segmentation on an abstract probability signal represented by signal p[1] . . . p[N].

According to an embodiment of the present invention, the size of the problem can be reduced by grouping alphabet symbols together in larger sets, called grid cells. See also the section “Fast Approximations” of Annex A, pages A2-A3, where N indicates the number of alphabet symbols, and Ñ indicates the number of grid cells. The size of the problem is reduced at the cost of approximate results. True optimality is obtained when N=Ñ.

1. Single-Encoder Quantization

Fixed rate quantization has already been defined in the introductory portion of the present application. Fixed rate problems are currently being solved by use of a source coding paradigm, as shown, for example, in references [1], [2], [3] and [4] of Annex A.

The present invention improves the prior art by providing a method which applies a variation of the traditional source coding paradigm also to variable rate scenarios, thus extending application also to environments where the number of the approximating symbols of the quantizer is not fixed beforehand and descriptions of different lengths are allowed, allowing application to a broader set of communication systems and better performance.

In a first step, the target of the optimization is defined. Such target is a function to be minimized, as shown by the Lagrangian J in equation (1) of the section “Optimization Criterion” of Annex A, page A3. Minimization of the Lagrangian J will allow the best quantizer to be obtained. Rate and distortion are broken down into a sum of per-codegroup partial rates and partial distortions, in order to transform the optimal quantization problem, i.e. the minimization of the Lagrangian J, into a graph problem, as shown in the same section of Annex A.

As already noted above, the use of a Lagrangian multiplier introduces a rate/distortion tradeoff on which the minimization problem according to the present invention is based. In the prior art, a tradeoff between rate and distortion is not specified and cannot be determined. Rather, a bound on rate or distortion is given. This bound is usually called “rate budget” and defines the maximum available rate or alternatively the maximum acceptable distortion.

Rate/distortion problems are presently very common, because of the explosion of the digital media. For example, when saving an image, there is usually a choice of low, medium or high compression. The higher the compression, the lower the rate (i.e. the file size) and the higher the distortion. Another example is given by cell phone technology: with a large bandwidth, more “rate” can be used and the speech sounds better. With a short amount of bandwidth, “low-rate” codes have to be used, which approximate the voice signal in order to send less information (bits) per unit time. In a storage scenario, the rate is measured in bits/sample. A sample could be, for example, a pixel of an image. For uncompressed black-and-white images (fax images), one bit per pixel would be needed (1=black, 0=white). For uncompressed grayscale images, 8 bits/pixel would be needed (from 0=darkest to 255=lightest). For a sound file, a sample would be the instantaneous level of the waveform. For example, a triangular waveform would have samples 0,1,2,0,1,2 . . . etc. In a transmission scenario, the rate is also measured in bit/sample, but sometimes also a bit/second measurement is used.

In the minimization step according to the present invention, various multipliers are tried, so that various rate/distortion combinations are generated, until the best rate/distortion combination is determined for a particular application.

In a second step, a relation between the problem of signal segmentation and the single-source shortest path problem in a weighted directed acyclic graph (WDAG) is obtained. Such step is known as such. See, for example, reference [6]. In particular, given the probability histogram p[1] . . . p[N] and a predetermined cost measure, the optimal quantization problem can be cast as an optimal segmentation problem on the histogram p[1] . . . p[N]. Equivalence between optimal segmentation problems and the shortest-path graph is known as such, as also shown in references [7]-[11].

In a third step, a minimization algorithm is applied, corresponding to the optimal segmentation algorithm. The optimal segmentation algorithm according to the present invention is equivalent to the shortest-path algorithm on a WADG called the “partial rate-distortion (RD) graph,” as explained in the section “Segmentation and Shortest Path Problems” of Annex A, pages A3 and A4. In particular, for every pair (u,v) with u<v there is an arc (directed edge) from u to v. The arc corresponds to a codegroup comprising cells u+1 . . . v and the weight is the Lagrangian cost d(C u+1 . . . v )+λr(C u+1 . . . v ). In the same section, the duality between edge and codegroups, and consequently between paths and partitions, is explained. This follows from the fact that paths are made up of edges and partitions are made up of codegroups.

The section “Algorithms” of Annex A, pages A4 and A5, describes the shortest-path algorithm and the topological sort algorithm which are applied to minimize the Lagrangian J. Such algorithms are known as such, and do not need to be described more in detail than already done at pages A4 and A5. See, for example, reference [12].

2. Multiple-Decoder Systems

The above embodiment can also be generalized to source codes designed for more general network scenarios.

In multiple decoder systems several descriptions are sent independently, and each description may or may not be received. Examples of single-source, multiple-decoder systems include multiple description scalar quantizers and multi-resolution scalar quantizers.

In multiple description scalar quantization (MDSQ), the source description is broken into a collection of packets, and reconstruction under any combination of received and lost packets is required. MDSQ is useful in diversity systems, where the encoder sends separate descriptions of the source over M different channels and reconstruction is required using the successfully received descriptions. MDSQ is also useful for packet-based communication systems and distributed data storage systems, where reconstruction of the data using a subset of the original set of source descriptions may be required.

Multi-resolution scalar quantization (MRSQ) is useful in applications where the compressed data sequence is intended for multiple users with differing rate and reproduction quality needs. In MRSQ, embedded source descriptions are used. The embedded source description may be decoded in part or in whole, with an expected reproduction quality that improves as a function of the rate of the decoded sequence. Since the binary description may be decoded at a variety of rates, the MRSQ decoder effectively contains a collection of decoders, one for each rate at which the binary sequence may be decoded.

MRSQ is a special case of MDSQ where decoders are designed for a subset of the possible packet-loss scenarios rather than all packet-loss scenarios. Therefore, an optimal MDSQ design algorithm leads to an optimal MRSQ design algorithm.

In the following, the description will initially focus on variable-rate MDSQ design with M=2, which will be indicated as 2DSQ. The M>2 scenario will then follow.

2A. Set-Up of the 2DSQ Method

The encoder of a 2DSQ gives two separate data descriptions and is therefore defined by two distinct partitions, P1 and P2, of the alphabet. The two descriptions can either be received alone or simultaneously. If the two descriptions are received alone, the problem reduces to ordinary single-description (SD) quantization with partition P1 and P2. Should the descriptions be received simultaneously, a new partition P0, called central partition, is defined, which contains codegroups pertaining both to the first and the second partition P1, P2, called side partitions.

Based on the central partition P0 and the side partitions P1 and P2, a target function is defined, which will be used as the optimization criterion for partition design in variable-rate 2DSQs. For a more detailed disclosure of the target function, see the section “Set-Up” of Annex A, pages A5 and A6.

2B. Solution of the 2DSQ Method

The method according to the present invention jointly optimizes the two side partitions P1 and P2, also keeping track of the central partition P0, which is a dependent function of the two side partitions P1 and P2.

First of all, a partition-set interpretation of the vertices in a 2DSQ scenario is provided. In a one-dimensional scalar quantization (SQ), each vertex represents a set of partitions that spans the cell range 0-v. In 2DSQ, vertex labels are pairs (vi, vj) and vertex (v1, v2) represents a set of partition pairs (P1, P2) such that P1 spans the range 0-v1 and P2 spans the range 0-v2.

When building the edge set, only one side partition may be extended at a time, thus only one side codegroup is added by each arc. Therefore, in principle, there should be an arc from vertex (v1, v2) to all vertices (v1′, v2) and (v1, v2′) such that v1′>v1 and v2′>v2. The difficulty arises when we attempt to assign weights to the edges. Some partitions must not be represented in their associated vertex and to guarantee that exclusion, some edges must be removed from the graph.

Computation of the central partition P0 is straightforward for side partitions having equal length, i.e. vertices (v, v). See FIG. 3 of Annex A and related discussion in Annex A. However, when the side partitions have different lengths, the central partition can be determined only up to the end of the shorter side partition. See FIGS. 5 and 6 of Annex A, and relative explanation at pages A6 and A7.

The consequence of ambiguity in the central partition definition is a corresponding difficulty in assigning weights to the arcs of the graph. In particular, while each arc corresponds to a single codegroup being added to one of the side partitions, the same arc corresponds to one or possibly multiple codegroups being added to the central partition. To be able to assign a unique weight to that arc, the total Lagrangian cost of the newly added codegroups must be the same regardless of the “internal structure” of the partition being extended. That is, the Lagrangian cost must be a function only of the labels of the two vertices that it joins. The codegroup added to one of the side partitions is clearly determined by the two vertex labels only, but the central codegroups are not, in general.

The inventors have successfully noted that the problem seems to arise from the fact that one of the side groups is “too short” relative to the other. Therefore, the inventors have developed a method to keep side partitions more “even” to avoid this type of problem. In particular, a “codegroup lag” has been defined, where partition P1 “lags” partition P2 by L codegroups if there are L codegroups on P2 not lying entirely inside the range covered by P1.

Therefore, the method according to the present invention keeps the codegroup lag between the two side partitions at most 1. A lag of zero will mean that the partitions are of equal total length. A partition pair having a codegroup lag less than or equal to 1 will be called valid.

Once the weight on each arc is defined as the Lagrangian cost associated with one new side codegroup and one new central codegroup, the above condition will allow a single weight to be defined for each arc.

The validity condition guarantees that adding side codegroups results in predictable central codegroups being generated. The following step of the method according to the present invention will be that of constraining the partial RD graph to obey the validity condition.

The method according to the present invention keeps partition pairs with codegroup lag more than 1 unrepresented in the vertices of the graph. Since the partition pairs represented in each vertex are defined by the paths between (0,0) and that vertex, removing arcs from the graph will be enough to eliminate the undesired paths.

The Partition Extension Rule

FIG. 5 of the present application (corresponding to FIG. 4 of Annex A) shows the partition extension rule used in solving the 2DSQ method. When v1 differs from v2, only the shorter side partition will be extended. When v1 equals v2 one of the two side partitions will be extended, for example P1. In this way some of the vertices are excluded. With reference to FIG. 5 :

in vertex (0, 0) v1=v2. Therefore, arcs connecting (0,0) with (1,0) and (2,0) will be generated;

in vertex (1,0) v2 (=0) is shorter than v1(=1). Therefore, arcs connecting (1,0) with (1, 1) and (1, 2) are generated;

in vertex (1, 1) v1=v2. Therefore, an arc connecting (1,1) with (2,1) is generated;

in vertex (2,1) v2(=1) is shorter than v1(=2). Therefore an arc connecting (2,1) with (2,2) is generated;

in vertex (1,2) v1(=1) is shorter than v2(=2). Therefore an arc connecting (1, 2) with (2,2) is generated; and

in vertex (2,0) v2(=0) is shorter than v1(=2). Therefore, arcs connecting (2,0) with (2,1) and (2,2) are generated.

It follows that some arcs are eliminated and some vertices become inaccessible from (0,0) so that they cannot be visited. In general, all vertices of the form (0,v2) with v2>0 are excluded. In FIG. 5 , vertices (0,1) and (0,2) are excluded.

The next step will be that of assigning a weight to each arc from (v1, v2) to (v′1, v′2). The weight on each arc is defined as the Lagrangian cost associated with one new side codegroup and sometimes one new central codegroup as explained at page A7 of Annex A. This is done according to the following criteria:

when v1=v2=v′2 (vertical arc starting from partitions having the same length) an additional codegroup is added to P1 and no codegroups are added to P0;

when v2=v′2>v1 (vertical arc starting from partitions having different lengths) an additional codegroup is added to P1 and one codegroup is added to P0; and

when v1=v′1>v2 (horizontal arc starting from partitions having different lengths) an additional codegroup is added to P2 and one codegroup is added to P0.

In this way, a partial RD graph has been built for the 2DSQ case. The subsequent step will be that of running one of the prior art shortest path algorithms described in the sections above, which will yield the optimal 2DSQ with contiguous codecells.

2C. Extension from 2DSQ to MDSQ

In the following, 2SDQ will be generalized by allowing M>2 quantizers to be designed jointly for M separate descriptions that travel independently. An MDSQ with M descriptions may be described by M side partitions which induce 2 M −M−1 non-trivial intersection partitions. The intersection partitions are the analogues of the central partition P0 from the 2DSQ scenario.

In this section, it will be assumed that the packet-loss scenarios corresponding to all of these intersection partitions occur with nonzero probability and thus that all are constrained. That assumption will be relaxed later.

Each possible packet-loss scenario can be described by a binary vector of length M, called the packet-loss vector; in this vector, bit i is 1 if packet i is received and 0 if packet i is lost in the corresponding scenario. Therefore, a packet-loss vector encodes information about a particular combination of lost and received descriptions. For example, packet-loss vector (1101) describes the scenario where, of an initial set of 4 packets, all but packet 3 are received. Side and intersection partitions can be indexed according to the same scheme. For example, P(0010) is the side partition used to describe packet 3 out of 4, while P1101 is the intersection partition corresponding to combined knowledge of packets 1, 2, and 4.

Also in this case a target Lagrangian function to be minimized, is described, as shown in Annex A, section “Extension from 2DSQ to MDSQ”, pages A8 and A9. In particular, the target is redefined to balance the performance of the side partitions and the intersection partitions. In the same section the graph construction is generalized by generalizing the partition extension rule and the arc weight definitions. Here, vertices are not just pairs, but M-tuples of integers since they must keep track of the growth of M side partitions. In the present scenario, the partition extension rule specifies which partition may be grown for any M-tuple, and also identifying a criterion for “breaking ties”, where the competitor with the smallest index is chosen.

As shown in Annex A, the details of implementing the shortest path algorithm for these particular graph structures are worked out. In case an ancestor-based version of the shortest path algorithm is implemented, as shown in the Annex A, the ancestor set of each vertex for the ancestor-based shortest path algorithm is explicitly computed. Should other versions of the shortest-path algorithm be used, the ancestor set is not required.

Also in this case, the method has a polynomial complexity as established in Annex A, page A9.

2D. Fixed-Rate MDSQ

In fixed-rate MDSQ design, the sizes of the side partitions are fixed and the rate of each description equals the logarithm of the corresponding size of the side partition. The generalization of the above method to fixed-rate MDSQ is described in the section “Fixed-Rate MDSQ” of Annex A, at page A9. In particular, the graph structure is refined to solve the fixed-rate version of the optimal quantization. In the present scenario, the vertices of the graph will also keep a count of how many segments each partition contains, so that a “double labeling,” one for the partition boundaries and one for the segments contained, is employed. In particular, each vertex is labeled with a tuple of M pairs of integers, for a total of 2M integers. Since rate is fixed, the Lagrangian cost incorporates only distortions.

In the following ‘MRSQ’ and ‘Restricted MDSQ’ sections, the admissible scenario set will be introduced, namely the set of possible combinations of received/lost descriptions. If some descriptions cannot be received without others, they may be encoded with less rate, so that the objective function for the optimization changes and new algorithms are employed.

2E. MRSQ

An overview of the MRSQ and a definition of the target function to be minimized is given in the section “MRSQ Overview” of Annex A, at page A10. In particular, multi-resolution coding is introduced and rate requirements are computed. In this case, a recursion relation for the target function is developed, which relation leads to a specialized MRSQ algorithm.

A detailed description of the method is given in the section “Algorithm” of Annex A, at pages A11 through A13. In particular, the recursion relation for the target function is exploited to formulate a multi-level dynamic programming algorithm. The algorithm starts at the lowest resolution level and builds a new set of weights at each level based on running an “all-pairs shortest-path” on the weights from the level directly below.

2F. Restricted MDSQ

Here, the most general case is treated, where a flexible set of dependencies can operate on the M descriptions. For example, a set of dependencies could be “description 3 cannot be received without 1 and 2, and description 5 cannot be received without 2 and 3.” A simple application implementing the concepts of the present section could provide, for example, a combination of two independent multi-resolution channels, like a 2DSQ using one 2RSQ and one 3RSQ. Three steps are followed:

1] Dependencies are extracted from the admissible scenario set;

2] Indirect dependencies are eliminated. For example, if 3 depends on 2 and 1, and 2 depends on 1, then the dependency of 3 on 1 can be eliminated.

3] The descriptions are then renumbered by sorting them topologically, according to the dependency graph.

With reference to dependencies and refinement, it is also noted that dependency between descriptions/packets does not necessarily imply refinement between corresponding side partitions, as shown at page A12 of Annex A. However, the inventors have proved that imposing refinement relationships between side-partitions corresponding to dependent descriptions causes no loss of generality: for any set of non-complying side partitions P′, another set of side partitions P can be constructed such that

1] The sets of intersection partitions generated by P and P′ are identical on the admissible scenario set;

2] P complies with the refinement requirement for side-partitions corresponding to dependent descriptions.

Once it is ensured that dependents are never extended beyond the boundaries of their requisites, refinement is preserved. This modification of the extension rule can be implemented directly also with the available general-MDSQ partial RD graph structure.

However, the partial RD graph structure has to be changed in order to assign weights to edges involving extensions of dependent descriptions. In order to compute the partial rate for a codegroup in a dependent partition, the “parent” codegroup in the requisite partition must be known. The end of that parent codegroup is recorded in the graph structure, but the beginning of the parent codegroup is not recorded yet. Thus, the vertices of the graph must be “double-labeled” to include the beginning, as well as the end, of codegroups in requisite partitions. If fixed-rate is also desired, this amounts to “triple-labeling” of the vertices: beginning of last codegroup, end of last codegroup, and number of codegroups. For partitions which are not requisites of any other partitions, this information is not necessary. New codegroup weights for dependent descriptions, which make use of the extra information available in requisite partitions, are derived. Because dependent packets are always received in the company of their requisites, more efficient encoding is possible.

Restricted-MDSQ can be combined with fixed-rate. It can be noted that any and all partitions may be designated as either fixed or variable rate. Thus, single-labeled, double-labeled and triple-labeled partitions may be present in the same graph.

3. Decoder Side-Information

The section ‘Decoder side-information’ (SECSQ) of Annex A, pages A13 and A14, treats the case when the decoder or receiver has access to both

A] what was transmitted by the encoder or sender and

B] a random variable Y correlated with the source X.

Of course, the receiver does not have access to X directly, as shown in FIG. 6 . There are two independent paths through which the source message X is communicated to the decoder: the encoding of X provided by the sender, and the message Y which is correlated with X. The decoder knows the joint distribution of X and Y, i.e. the individual probability that X=x and Y=y for all pairs <x,y>.

For example, suppose X is the temperature in New York (in Fahrenheit, with 1-degree precision), the sender is located in New York, and the receiver is in Los Angeles. The sender describes X to the receiver using some quantizer, e.g. by using 10 degree increments: if X is between 50 and 60, the sender encodes it as 55; if X is between 60 and 70, the sender encodes it as 65 etc. Further suppose that, for some reason, the decoder has access to Y, the last digit of X (3 for 63, for example). By combining knowledge from the received approximation of X and from Y, the decoder can tell exactly what X was (most real-world examples would not be as dramatic). In this example, the joint probability for <x, y> is simply 1 if y is the last digit of x and 0 otherwise. The source alphabet is 0 . . . 100, and the side-information alphabet is 0 . . . 9.

At pages A13 and A14 of Annex A, the inventors prove that in the case of SECSQ (ECSQ with side information) the same partial RD graph structure can be used, but with modified weights, to obtain the optimal quantizer. Annex A also shows how speed-up techniques used to reduce the computational complexity of determining ECSQ weights translate directly to the SECSQ problem, the only modification being that the preprocessing techniques must be repeated N(Y) times, where N(Y) is the size of the side-information alphabet.

Furthermore, for multiple-decoder with side information systems, the same pattern applies: keep the graph structure from no-side-information versions, but use the partial RD formulas and speed-up techniques developed for SECSQ.

4. Codecell Contiguity

Pages A14 to A17 of Annex A discuss whether and when the quantizer regularity restriction (i.e. the requirement of codecell contiguity) precludes true optimality. The following conclusions are reached:

1] Fixed-rate optimal ECSQs (entropy-constrained scalar quantizers) always use contiguous codecells for distortion measures that are non-decreasing functions of the absolute error, see Theorem 1 at page A14 of Annex A. Also the mean-squared error (MSE) is included in this class of distortion measures.

2] For variable-rate ECSQs, two further constraints must be imposed to guarantee contiguous codecells, see Theorem 2 at page A14 of Annex A:

a] The distortion measure must be a convex function of the absolute error, not just non-decreasing. MSE satisfies this condition.

b] Contiguity is claimed only for points on the lower convex hull of the rate-distortion curve. However, this kind of constraint is somewhat academic, because Lagrangian-based optimization only sees points on the convex hull anyway.

3] Fixed- and variable-rate MRSQ, MDSQ (general or restricted) and SECSQ (side-information ECSQ) may require non-contiguous quantizers, regardless of the distortion measure. See also Theorem 5 at page A16 of Annex A.

4] For MRSQ with the MSE distortion measure, while the finest partition has to be contiguous in an optimal quantizer, the coarsest partition can be non-contiguous.

While several illustrative embodiments of the invention have been shown and described in the above description and in the enclosed Annex A, numerous variations and alternative embodiments will occur to those skilled in the art. Such variations and alternative embodiments are contemplated, and can be made without departing from the scope of the invention as defined in the appended claims.

Claims

1. A signal encoding system having a coder for encoding signals, the coder including a scalar quantizer and a lossless coder, the scalar quantizer producing a distortion on the signals to be encoded, wherein the scalar quantizer is globally optimal quantizer under a rate/distortion tradeoff, so that the distortion is minimized when the rate of the optimal quantizer is no larger than a specified value.

2. The signal encoding system of claim 1, wherein the scalar quantizer is a variable rate quantizer.

3. The signal encoding system of claim 1, wherein the signal encoding system is a communications system and the encoded signals are transmitted along a channel.

4. The signal encoding system of claim 1, further comprising a storage device for storing the encoded signals.

5. The signal encoding system of claim 1, wherein the scalar quantizer is a single-description scalar quantizer.

6. The signal encoding system of claim 1, wherein the scalar quantizer is a multiple-description scalar quantizer.

7. The signal encoding system of claim 6, wherein the scalar quantizer is a multi-resolution scalar quantizer.

8. A signal decoding system having a decoder for decoding signals, the decoder including a lossless decoder and an inverse scalar quantizer wherein the inverse scalar quantizer is globally optimal quantizer under a rate/distortion tradeoff, so that the distortion is minimized when the rate of the optimal quantizer is no larger than a specified value.

9. The signal decoding system of claim 8, wherein the inverse scalar quantizer is a variable rate inverse scalar quantizer.

10. The signal decoding system of claim 8, wherein the data decoding system is a communications system and the decoded signals are received through a channel.

11. The signal decoding system of claim 8, further comprising a storage device for storing the decoded signals.

12. The signal decoding system of claim 8, wherein the inverse scalar quantizer is a single-description scalar quantizer.

13. The signal decoding system of claim 8, wherein the inverse scalar quantizer is a multiple-description scalar quantizer.

14. The signal decoding system of claim 13, wherein the inverse scalar quantizer is a multi-resolution scalar quantizer.

15. A signal encoding method for use in a signal encoding system, comprising the steps of:providing a source alphabet containing a set of source characters; approximating the set of source characters by designing a scalar quantizer and applying the designed scalar quantizer to the set of source characters, thus producing a smaller set of characters and a distortion on the set of source characters; and applying a compression algorithm to the smaller set of characters, wherein the step of designing a scalar quantizer comprises the steps of: defining a target function depending on the distortion on the set of source characters; and minimizing the target function to obtain a globally optimal quantizer.

16. The method of claim 15, wherein the target function also depends on the number of bits describing a source character in the encoding method.

17. The method of claim 16, wherein the minimized target function is optimal under a rate/distortion tradeoff, the rate being the number of bits describing a source character in the encoding method.

18. The method of claim 17, wherein the step of minimizing the target function comprises the steps of:dividing the rate into a sum of partial rates; and dividing the distortion into a sum of partial distortions.

19. The method of claim 15, wherein the step of minimizing the target function comprises the step of building a graph.

20. The method of claim 19, wherein the target function is minimized by applying a shortest path algorithm to the graph.

21. The method of claim 20, wherein the shortest path algorithm is a single-source shortest path algorithm.

22. The method of claim 20, wherein the shortest path algorithm is an all-pairs shortest path algorithm.

23. The method of claim 20, wherein the shortest path algorithm is an optimal segmentation algorithm.

24. The method of claim 20, wherein the step of applying the shortest path algorithm is applied to a partial rate-distortion graph.

25. The method of claim 19, wherein the graph is a weighted directed acyclic graph.

26. The method of claim 15, wherein the step of approximating the set of source characters comprises the step of obtaining a partition of the source alphabet.

27. The method of claim 15, further comprising a step of preprocessing the source alphabet before the step of approximating the source alphabet.

28. The method of claim 15, wherein the target function is minimized by means of dynamic programming.

29. A client-server system comprising a server and a plurality of clients, wherein when a client of the plurality of clients requests data to the server, the server provides the client with an approximation of the requested data, the approximation being based on a quantization process of the requested data, thus producing distortion on the requested data, the quantization process approximating the requested data to a smaller number of data and comprising a step of designing a scalar quantizer by defining a target function and minimizing the target function to obtain a globally optimal quantizer.

30. The system of claim 29, wherein the minimized target function is optimal under a rate/distortion tradeoff, the rate depending on the smaller number of data.

31. The system of claim 29, wherein the step of minimizing the target function comprises the step of building a graph.

32. The system of claim 31, wherein the target function is minimized by applying a shortest path algorithm to the graph.

33. The system of claim 31, wherein the step of minimizing the target function is performed by means of dynamic programming.