RU2426227C2 - Encoding variable-length codes with efficient memory usage - Google Patents

Abstract

FIELD: information technology.

SUBSTANCE: invention is aimed at methods for adaptive variable-length coding (VLC) with efficient memory usage and low degree of complexity with respect to data for different fields of application, such as digital video data, image data, audio and voice data. In certain aspects, these methods can employ properties of defined sets of code words to support very compact data structures. In other aspects, the methods can support adaptive coding and decoding with low degree of complexity with respect to binary sequences generated by memoryless sources.

EFFECT: adaptive variable-length coding with more efficient memory usage and low complexity for data in different fields of application, such as coding digital video data, image data, audio or voice data.

83 cl, 15 dwg

Description

This application claims priority on the basis of provisional application US registration No. 60 / 865,827, filed November 14, 2006, and provisional US application registration number 60 / 867,081, filed November 22, 2006, the contents of each of which is fully incorporated into this application by links.

FIELD OF THE INVENTION

The present invention relates to data compression and, more specifically, to data compression using variable length codes (VLC).

State of the art

Data compression is widely used in various applications to reduce the occupied area for data storage, bandwidth, or both. Examples of data compression applications include encoding digital video data, image data, voice and audio data. Digital video encoding is used, for example, in a number of devices, including digital television devices, digital direct broadcasting systems, wireless communication devices, personal digital assistants (PDAs), laptop computers or desktop computers, digital cameras, digital recording devices, devices for video games, cellular or satellite radiotelephones and the like. Digital video devices implement video compression methods, such as MPEG-2, MPEG-4, or H.264 / MPEG-4 enhanced video coding (AVC), for more efficient transmission and reception of digital video data.

In general, in video compression methods to reduce or eliminate redundancy inherent in video data, spatial prediction, motion estimation, and motion compensation are performed. In particular, intra-coding relies on spatial prediction to reduce or eliminate spatial redundancy in video data within a given video frame. External coding relies on temporal prediction to reduce or eliminate temporal redundancy in video data within adjacent frames. For external coding, the video encoding device performs motion estimation to track the movement of the respective video blocks between two or more adjacent frames. The motion estimation generates motion vectors that show the offset of the video blocks relative to the corresponding video blocks in one or more reference frames. Motion compensation uses a motion vector to generate a prediction video block from a reference frame. After motion compensation by subtracting the prediction video block from the original video block, a residual video block is formed.

The video encoding apparatus employs transform encoding, quantization encoding, and statistical encoding processes to further reduce the bit rate of the residual block obtained in the video encoding process. Statistical encoding methods are used on the last stages of the video signal encoder / decoder (CODEC) and various other encoding applications before storing or transmitting the encoded data. Statistical coding as a whole involves the use of arithmetic codes or variable length codes (VLCs) to further compress the residual coefficients resulting from the transform and quantization operations. Examples of statistical coding methods include context-sensitive binary arithmetic coding (CABAC) and variable-length context-sensitive coding (CAVLC), which can be used in some encoders as alternative statistical coding modes. The video decoding apparatus performs statistical decoding to decompress residual information for each of the blocks and recover encoded video data using motion information and residual information.

SUMMARY OF THE INVENTION

In General, the present invention is directed to the creation of adaptive coding of variable length (VLC) with more efficient use of memory and less complexity for data in various applications, such as encoding digital video data, image data, audio or voice data. In a first general aspect, certain codeword sets can be used in these methods to support very compact data structures. In a second general aspect, the methods may support adaptive coding and decoding with a low degree of complexity of binary sequences generated by memoryless sources.

The invention provides, in a first aspect, a method comprising generating partial values of basic codewords for levels of a coding tree defining variable-length codewords, generating a skip indicator instructing the decoder to skip a few bits in the bitstream to be decoded before proceeding to the selected level of the coding tree , and storing partial values and skip in the data structure in memory.

In another aspect, the invention provides a tangible computer-readable medium comprising a data structure storing partial values of base codewords for levels of a coding tree defining variable-length codewords and a skip indicator instructing the decoder to skip a few bits in the bitstream to be decoded before proceeding to the selected coding tree level.

In an additional aspect, the invention provides a device comprising a processor configured to generate partial values of basic codewords for levels of a coding tree defining variable-length codewords and generating a skip indicator instructing the decoder to skip a few bits in the bitstream to be decoded before proceeding to the selected the level of the coding tree, and a memory that stores partial values and a skip rate in the data structure.

In yet another aspect, the invention provides a decoding apparatus comprising a memory storing a data structure comprising partial values of base codewords for levels of a coding tree defining variable-length codewords, and a skip indicator instructing the decoder to skip several bits in the bitstream to be decoded before a transition to the selected level of the encoding tree, and a decoder that accesses the memory to decode one of the code words from the bitstream based on partial values and a skip rate in the stored data structure.

In yet another aspect, the invention provides a decoding method comprising accessing a data structure stored in a memory, the data structure comprising partial values of base codewords for levels of a coding tree defining variable length codewords, and a skip rate instructing the decoder to skip a few bits in the stream the bits to be decoded before moving to the selected level of the encoding tree, and decoding one of the code words from the bitstream based on partial values and exponent pass in the stored data structure.

In yet another aspect, the invention provides a tangible computer-readable medium comprising instructions causing a processor to access a data structure stored in memory, the data structure comprising partial values of base codewords for levels of a coding tree defining variable-length codewords and a skip indicator indicating a decoder skip the many bits in the bit stream to be decoded before moving to the selected level of the encoding tree, and decode one of the code words from n bits of the current based on the partial values and the skip indicator in the stored data structure.

In an additional aspect, the invention provides a method comprising performing variable-length coding according to a code structure, the code structure defining groups of codewords in a coding tree, each of the groups includes codewords representing values having the same weights, and codewords in each of the groups ordered in lexicographic order with respect to the values represented by the code words, and the first and second subgroups of code words within each of the groups, the first subgroup including by codewords having a first length, and the second subgroup includes codewords having a second length different from the first length, and generating a variable length encoding result for at least one of being stored in memory, transmitted to a device, or presented to a user.

In an additional aspect, the invention provides a tangible computer-readable medium comprising instructions causing a processor to perform variable-length coding according to a code structure, the code structure defining groups of code words in a coding tree, each group including code words representing values having the same weights, and the code words in each of the groups are ordered in lexicographic order relative to the values represented by the code words, and the first and second subgroups of code words within each of the groups, wherein the first subgroup includes codewords having a first length, and the second subgroup includes codewords having a second length different from the first length, and generate a variable length encoding result for at least one of the storage in memory, transferring to a device or presentation to a user.

In an additional aspect, the invention provides a device comprising a processor configured to perform variable length coding according to a code structure, the code structure defining groups of code words in a coding tree, each of the groups includes code words representing values having the same weights and code words in each of the groups are ordered in lexicographic order relative to the values represented by code words, and the first and second subgroups of code words inside each from groups, the first subgroup including code words having a first length, and the second subgroup includes code words having a second length different from the first length, and generating a variable length encoding result for at least one of the memory, transmitting to the device or presentation to the user.

In a further aspect, the invention provides a method comprising for a code structure definition of a group of codewords in a coding tree, determining variable length codewords, wherein each of the groups includes codewords representing values having same weights and the codewords in each of the groups are ordered in lexicographic order with respect to the values represented by the code words, the first and second subgroups of code words within each of the groups, the first subgroup including code layers va having a first length and the second subgroup includes codewords having a second length different from the first length, performing variable length coding using base codewords for each of the subgroups, positions of codewords within each of the groups, the number of code words in each of the first subgroups and codeword lengths in each of the subgroups, and generating a variable length encoding result for at least one of being stored in memory, transmitted to a device, or presented to a user.

In an additional aspect, the invention provides an apparatus comprising, for code structure, means for determining groups of codewords in a coding tree defining codewords by varying lengths, each of the groups including codewords representing values having the same weights and codewords in each of the groups are ordered in lexicographic order relative to the values represented by the code words, and the first and second subgroups of code words within each of the groups, the first subgroup includes common words having a first length, and the second subgroup includes code words having a second length different from the first length, a means of performing variable-length encoding using basic code words for each of the subgroups, positions of code words in each of the groups, the number of code words in each of the first subgroups and codeword lengths in each of the subgroups, and means for generating a variable length encoding result for at least one of being stored in memory, transmitted to a device, or presented to a user.

In an additional aspect, the invention provides a tangible computer-readable medium comprising instructions causing a processor for a code structure defining groups of codewords in a coding tree, each of which includes codewords representing values having the same weights and codewords in each of groups are ordered in lexicographic order relative to the values represented by the code words, and the first and second subgroups of code words within each of the groups, the first subgroup including includes code words having a first length, and the second subgroup includes code words having a second length different from the first length, perform variable length encoding using basic code words for each of the subgroups, codeword positions in each of the groups, the number of codewords in each of the first subgroups and the lengths of codewords in each of the subgroups, and generate a variable length encoding result for at least one of being stored in memory, transmitted to a device, or presented to a user.

In an additional aspect, the invention provides an apparatus comprising, for a code structure defining groups of codewords in a coding tree, each of the groups includes codewords representing values having the same weights, and codewords in each of the groups are ordered in lexicographic order relative to the values represented by code words, and the first and second subgroups of code words within each of the groups, the first subgroup includes code words having a first length and a second subgroup includes codewords having a second length different from the first length, a processor configured to perform variable length coding using basic codewords for each of the subgroups, positions of the codewords in each of the groups, the number of codewords in each of the first subgroups and codeword lengths in each of the subgroups, and generating a variable length encoding result for at least one of being stored in memory, transmitted to a device, or presented to a user.

In yet another aspect, the invention provides a tangible computer-readable medium comprising a data structure for variable-length encoding using a variable-length code structure that defines groups of codewords in a coding tree, each of the groups includes codewords representing values having the same weights, and the code words in each of the groups are ordered in lexicographic order relative to the values represented by the code words, and the first and second subgroups of code words inside each second groups, wherein the first subgroup includes codewords having a first length and the second subgroup includes codewords having a second length different from the first length.

In yet another aspect, the invention provides an integrated circuit device comprising a memory storing a data structure including partial values of base codewords for levels of a coding tree defining variable length codewords and a skip indicator instructing the decoder to skip a few bits in the bitstream to be decoded , before moving on to the selected level of the encoding tree, and a decoder that accesses the memory to decode one of the code words from the bitstream based on partial values and a skip rate in the stored data structure.

In yet another aspect, the invention provides a portable wireless communications device comprising a memory storing a data structure comprising partial values of base codewords for levels of a coding tree defining variable length codewords and a skip indicator instructing the decoder to skip multiple bits in the bitstream to be decoded , before moving on to the selected level of the encoding tree, a decoder that accesses the memory to decode one of the code words from the bitstream based on cha statistics and the skip rate in the stored data structure, a receiver for receiving codewords from the encoder via wireless communication, and an output device that presents the output signal to the user based at least in part on the decoded codewords.

In yet another aspect, the invention provides an integrated circuit device comprising a processor configured to perform variable length coding according to a code structure, the code structure defining groups of code words in a coding tree, each of the groups including code words representing values having the same the weights and codewords in each of the groups are ordered in lexicographic order relative to the values represented by the codewords, and the first and second subgroups of codewords s within each of the groups, the first subgroup includes codewords having a first length, and the second subgroup includes codewords having a second length different from the first length, and the processor is configured to generate a variable length encoding result from for at least one of storage, transmission to a device, or presentation to a user.

In yet another aspect, the invention provides a portable wireless communications device comprising a processor configured to perform variable length encoding according to a code structure, the code structure defining groups of code words in a coding tree, each of the groups including code words representing values having the same the weights and code words in each of the groups are ordered in lexicographic order relative to the values represented by the code words, and the first and second subgroups single words within each of the groups, the first subgroup including code words having a first length, and the second subgroup includes code words having a second length different from the first length, and the processor is configured to generate a variable length encoding result at least one of storage, transmission to a device, or presentation to a user.

The methods described in this application can be implemented in hardware, software, or a combination thereof. If implemented by software, the software may run on one or more processors, such as a microprocessor, application specific integrated circuit (ASIC), user-programmable logic array (FPGA) or digital signal processor (DSP), or other equivalent integrated or discrete logic circuit. Software that performs these methods may initially be stored in a computer-readable medium and loaded and executed by a processor. Accordingly, the present invention also provides computer program products comprising a computer-readable medium that comprises instructions for causing a processor to execute any of a variety of methods described herein.

Details of one or more aspects of the invention are provided in the accompanying drawings and in the following description. Other features, objectives and advantages of the methods described in this application will become apparent from the description and drawings, as well as from the formula.

Brief Description of the Drawings

Figure 1 is a block diagram illustrating a system for encoding and decoding video data.

2 is a block diagram illustrating an example of a video encoder.

3 is a block diagram illustrating an example of a video decoder.

4 is a diagram illustrating an example of a binary encoding tree.

5 is a graph illustrating a degree of redundancy of an adaptive block code with asymptotic behavior.

6 is a binary tree diagram illustrating block groups, subgroups, and basic codewords.

7A and 7B are graphs illustrating a comparison of the degrees of redundancy of an adaptive block code with various ρ values.

8 is a graph illustrating the sensitivity of redundancy to the asymmetry of the source data.

FIG. 9 is a flowchart illustrating a method for constructing variable length coding with efficient use of memory for monotone distributions in accordance with one aspect of the present invention.

FIG. 10 is a flowchart illustrating a method of encoding symbols using variable-length codes constructed according to the method of FIG. 9.

11 is a flowchart illustrating a method for decoding variable-length variable codes constructed according to the method of FIG. 9.

12 - a block flowchart illustrating a method for constructing adaptive block codes in accordance with another aspect of the present invention.

FIG. 13 is a flowchart illustrating a method of encoding blocks using variable-length codes constructed according to the method of FIG. 12.

FIG. 14 is a flowchart illustrating a method for decoding variable length codes constructed according to the method of FIG. 12.

Detailed description

In General, the present invention is directed to the creation of adaptive coding of variable length (VLC) with efficient use of memory and reduced complexity in relation to data for many applications, such as encoding digital video data, image data, audio or speech data. In some aspects, these methods may use the properties of certain sets of codewords to support very compact data structures. In other aspects, the methods may support low complexity, adaptive coding, and decoding of binary sequences generated by memoryless sources. Although the methods described in this application can be applied to a wide range of practical applications, including general compression and encoding of data, the encoding and decoding of digital video data is described in this application as an example and illustration.

In accordance with a first general aspect of the present invention, the disclosed methods do not need to rely on any particular code construction scheme to support compact data structures. For example, you can use Huffman, Shannon, Shannon-Fano, Gilbert-Moore, or other code construction schemes. However, for this first general aspect, it is assumed that such a code is constructed for a source with monotonically increasing probabilities of characters from the input character alphabet. More specifically, it is assumed that the length of the codewords is monotonically decreasing or at least not increasing, and that the codewords of the same length have the same lexicographic order as the characters in the input alphabet that they represent.

The desired lexicographical order can be provided by reordering the input alphabet. For such codewords, you can calculate the basic values of the codewords for each level of the coding tree. The base codeword values are the lexicographically smallest canonical codewords at each level of the coding tree. The base values of the codewords and the indices of their respective characters can be stored in a reordered array. These indices can be stored as offset values for each populated tree level. The decoding process may include comparing the bitstream buffer with the values of the left-aligned base codewords, followed by a simple direct calculation of the index of the decoded symbol.

The above properties can be used to uniquely describe such a code using a data structure that can be further compressed to obtain a very compact data structure that facilitates the phased decoding of variable-length codes. For example, left-aligned values of base codewords usually have a large number of leading zeros from right to left. The number of leading zeros decreases monotonically as the process moves to deeper layers in the corresponding coding tree. Therefore, when sequentially decoding bits, starting from the very first layer and then moving down, some of the initial zero bits can be skipped without affecting the accuracy of the decoding process.

Leading zeros can be deleted with fixed increments, for example, with 8-bit increments, which is convenient when controlling the data stream buffer in traditional 8-bit / byte computers. To control this process, an additional table may be provided containing one or more skip indicators. For example, a skip metric tells the decoder to scroll the bit stream forward in the buffer in a fixed increment so that different base codewords can be distinguished when the leading zeros are omitted. After removing the leading zeros, the width of the resulting array of modified values of the base codewords can be significantly reduced, thereby saving free memory.

Thus, in a first general aspect, the present invention provides a method comprising generating partial values of basic codewords for levels of a coding tree defining variable-length codewords, generating a skip indicator instructing the decoder to skip a few bits in the bitstream to be decoded before proceeding to the selected the level of the coding tree, and the preservation of partial values and the rate of omission in the data structure in memory. The data structure may be any of a variety of data structures, such as tables, linked lists, binary trees, base trees, unstructured files, or the like, and may be stored in any of a variety of different storage devices, such as various kinds of random access memory (RAM) , read only memory (ROM), or both. The data structure may be stored in such memory inside an encoder or decoder. For example, the decoder may access the data structure or portions of the contents of the data structure from the memory associated with the decoder to perform variable-length decoding for code words with efficient memory usage.

In accordance with this first general aspect, the invention further provides a processor configured to generate partial values of the base codewords for the levels of a coding tree defining variable length codewords, and generate a skip indicator instructing the decoder to skip a few bits in the bitstream to be decoded before a transition to the selected level of the encoding tree, as well as a memory that stores partial values and the skip rate in the data structure. Such information may be stored in a single data structure or in multiple data structures. Accordingly, the term “data structure” may refer to one or more data structures storing information provided by the present invention. A processor configured to access a data structure to perform variable-length encoding can be implemented in a source device or in a receiver device or in a separate device that generates data structures that define code structures for use by an encoder and / or decoder when performing encoding variable length.

According to this method of providing compact data structures in accordance with a first general aspect of the present invention, each valid codeword length may correspond to a level with an external node in the code tree. As indicated above, the data structure may include partial values of the base codewords and one or more missing indicators. More specifically, in some embodiments, the data structure may contain the following information for each valid codeword length: (a) the partial value of the lexicographically smallest (or largest) codeword at the current code tree level, (b) the number of bits in the partial value, (c ) the value of the character corresponding to the lexicographically smallest (or largest) codeword, and (d) an indicator that tells the decoder to skip a certain number of bits before moving on to the next level of the code tree. Accordingly, the data structure may further include values for the characters represented by the base codewords and the lengths of the partial values of the base core codewords, that is, the number of bits in each partial value of the base codeword. Encoding and decoding methods can use this data structure to identify the level corresponding to the codeword to be created or decoded, and then directly calculate the value of the codeword or decoded symbol. Accordingly, the data structure may be stored in the memory of a video encoder or decoder, an image encoder or decoder, an audio data encoder or decoder, or a voice data encoder or decoder, some of which may be designed as combined CODECs.

Examples of existing methods for encoding or decoding variable length codes are described in A. Moffat and A. Turpin, On the Implementation of Minimum-Redundancy Prefix Codes, IEEE Trans. Communications, 45 (10) (1997) 1200-1207; J. B. Connell, A Huffman-Shannon-Fano Code, Proc. IEEE, July 1973, 1046-1047; and A. Brodnik and S. Carlsson, Sublinear Decoding of Huffman Codes Almost in Place, DIMACS Workshop on Codes and Trees: Algorithmic and Information Theoretic Approaches, October 1998, Rutgers University, DIMACS Center, NJ. Compared to these existing methods, the disclosed methods for providing compact data structures may offer certain advantages.

As a first example, a data structure created by the disclosed method can occupy a significantly smaller amount of memory due to the fact that only partial values of the lexicographically smallest code words are stored and used, for example, by a video data decoder. As another example, the disclosed method may use step-by-step access to the bitstream data, which allows the bitstream buffer to be presented in a reasonably short register and updated at any convenient intervals, for example, by indicating a skip, which further reduces implementation complexity.

For example, in some embodiments, a 32-bit register may be sufficient even for very long codes. In addition, updates can be made at 8-bit intervals. In General, the disclosed method can significantly reduce the complexity of the presentation and encoding / decoding of codes of variable length, which may allow developers of compression algorithms to use significantly larger and, therefore, more efficient codebooks.

In accordance with a second general aspect of the present invention, to support adaptive coding and decoding with a low degree of complexity for binary sequences created by memoryless sources, the disclosed methods allow implementing universal block codes constructed for a set of contexts identified by the number of non-zero bits among previous bits in a sequence . This second general aspect can be provided and applied independently or in conjunction with the first general aspect described above with respect to the generation of very compact data structures. Methods of adaptive coding and decoding with a low degree of complexity can, in accordance with this second general aspect, rely on the disclosed formula for the asymptotic redundancy of such codes, which refines the estimate described by RE Krichevsky and VK Trofimov, The Performance of Universal Encoding, IEEE Trans . Information Theory, 27 (1981) 199-207.

The methods in accordance with this second general aspect can use an array of Huffman codes designed for several estimated densities and indexed by the number of nonzero bits in previous blocks (contexts) in the sequence. When using blocks of even moderate size n = 8 ... 16 (and using appropriate 1.5 ... 5 kilobytes of memory), such methods can provide compression efficiency comparable to or superior to other existing algorithms, such as the Q-coding algorithm described in W.B. Pennebaker, J.L. Mitchell, G.G. Langdon, Jr, R. B. Arps, An overview of the basic principles of the Q-coder adaptive binary arithmetic coder, IBM J. Res. Dev. 32 (6) (1988) 717, which is used in the JBIG image encoding standard, or the CABAC algorithm described by D. Marpe, H. Schwartz and T. Wiegand, Context-Based Adaptive Binary Arithmetic Coding in the H. 264 / AVC video compression standard, IEEE Trans. on CSVT, 13 (7): 620 636, July 2003, which is used in ITU-T H.264 / MPEG AVC standards for video compression.

Adaptive coding and decoding with a low degree of complexity, in accordance with this second general aspect of the invention, can be based on the understanding that in a memoryless model, the probability of a block of bits or its estimation depends only on its weight (the number of nonzero bits), but not on the actual combination of bits. Therefore, the set of all possible blocks of a certain fixed length can be divided into several groups containing blocks with the same weight (and, therefore, with the same probability). It can be assumed that each group stores the blocks in lexicographical order, either naturally or as a result of reordering.

It is a known property of codes with minimal redundancy (such as Huffman or Shannon codes) that each group of equally probable blocks can include a maximum of two subgroups, and the blocks in each such subgroup are encoded with code words of the same length. Without loss of generality, it can be further assumed that the codeword length in the first subgroup is shorter than the codeword length in the second subgroup. Since the blocks (or words) within the group follow in lexicographic order, it can simply be divided into subgroups with longer codewords and subgroups with shorter codewords. The position of the block within the group is indicated by the index value. Lexicographically, the smallest block (or word) in each subgroup is assigned a base codeword. With a basic codeword, you can easily calculate the rest of the codewords in each subgroup.

Thus, in accordance with this second general aspect of the invention, variable-length coding can be performed, for example, by an encoder or decoder using a code structure that defines groups of input blocks or words and their corresponding output code words in a coding tree, each group including code words representing blocks (or words) having the same weights and a first and second subgroup of code words within each of the groups, the first subgroup including code words having the first length, and the second subgroup includes codewords having a second length different from the first length. Blocks in each of the groups are ordered in lexicographic order and then divided into subgroups, so that the lexicographic order is preserved in each of the subgroups. In addition, codewords corresponding to each block within a subgroup are assigned in such a way that they also follow the same lexicographic order and facilitate encoding and decoding by direct calculation.

Based on this relative arrangement of blocks and their groups and subgroups, code words can be directly calculated using a simplified process. For example, after obtaining the weight and index value for a block, if the index value is less than the maximum number of blocks in the first subgroup, then the first subgroup is selected to determine the location of the code word. Otherwise, a second subgroup is selected to locate the codeword. Then, after extracting the base codeword for the selected subgroup, the codeword is easily calculated by summing the value of the base codeword with a value determined based on the value of the block index within the selected subgroup. With respect to this second general aspect of the invention, the terms “blocks” and “words” may be used interchangeably to generically denote input values to be encoded in accordance with a coding scheme. A block or word may include a sequence of characters that can be formed from an input alphabet, such as the binary alphabet {0, 1}. Code words generally denote the output values assigned to blocks (or words) as a result of applying the coding scheme.

These and other aspects of the methods described in this application are described in more detail below. The methods can be used together or independently and can be implemented in any of a variety of encoding and decoding systems and devices, including systems and devices configured to encode and decode digital video data, image data, audio or voice data. As an example and illustration, the application describes the application of such methods to the encoding and decoding of digital video data, which does not impose a restriction on the general practical application of data compression and encoding, or on other particular applications for various types of data.

1 is a block diagram illustrating a video encoding and decoding system 10. As shown in FIG. 1, system 10 includes a source device 12 that transmits encoded video data to a receiver device 14 via a communication channel 16. The source device 12 may include a video source 18, a video encoder 20, and a transmitter 22. A receiver device 14 may include a receiver 24, a video decoder 26, and a video display device 28. System 10 can be configured to employ adaptive variable length coding (VLC) methods with efficient memory use and low complexity for digital video data. In particular, adaptive VLC methods with efficient use of memory and low complexity can be used for statistical coding of residual block coefficients obtained in the process of encoding predictive video data. These methods can be applied to video encoding schemes that encode the positions of non-zero transform coefficients using sequences of zeros, or to other video encoding schemes.

In the example of FIG. 1, communication channel 16 may comprise any wireless or wired communication medium, such as a radio frequency (RF) spectrum or one or more physical transmission lines, or any combination of wireless and wired media. Channel 16 may be part of a packet-based network, such as a local area network, a regional network, or a wide area network, such as the Internet. The communication channel 16 as a whole represents any suitable communication medium or a combination of various communication media for transmitting video data from the source device 12 to the receiver device 14.

The source device 12 generates video data for transmission to the destination device 14. In some cases, however, devices 12, 14 may operate in a substantially symmetrical manner. For example, each of devices 12, 14 may include encoding and decoding components of video data. Therefore, system 10 can support one-way or two-way video transmission between video data devices 12, 14, for example, for streaming video, broadcasting video, or video telephony. For other applications, data compression and encoding of the device 12, 14 may be configured to send and receive, or exchange other types of data, such as image data, voice or audio data, or a combination of two or more of video data, image data, voice and sound data. Accordingly, consideration of the application with respect to video data is used for illustrative purposes and should not be construed as limiting the various aspects of the invention described herein in the broadest possible manner.

Video source 18 may include a video capture device, such as one or more video cameras, a video archive containing previously captured video images, or video data directly from a video content provider. As another alternative, the video source 18 can generate computer graphics based data or a combination of the direct video data and video data generated on a computer as source video data. In some cases, if the source of the video data 18 is a camera, the source device 12 and the receiver device 14 may form so-called video telephones. So, in some aspects, source device 12, receiver device 14, or both, can form a portable wireless device, such as a mobile phone. In each case, the captured, pre-captured, or computer-generated video image may be encoded by the video encoder 20 for transmission from the video source device 12 to the video decoder 26 of the video receiver device 14 of the video data through the transmitter 22, channel 16 and receiver 24. The display device 28 may include all kinds of display devices, such as a liquid crystal display (LCD), a plasma display, an organic light emitting diode (OLED) display.

Video encoder 20 and video decoder 26 may be configured to support scalable video encoding for spatial, temporal scalability and / or scalability of the signal-to-noise ratio (SNR). In some aspects, video encoder 20 and video decoder 22 may be configured to support fine mesh SNR scalability (FGS) coding for SVCs. Encoder 20 and decoder 26 can support varying degrees of scalability by supporting encoding, transmitting, and decoding of the base layer and one or more scalable enhancement layers. For scalable video coding base layer carries video data with a minimum level of quality. One or more layers carry additional quality improvement data flow to support higher spatial, temporal level and / or level of SNR.

Video encoder 20 and video decoder 26 may operate in accordance with a video compression standard such as MPEG-2, MPEG-4, ITU-T H.263 or ITU-T H.264 / MPEG-4 Advanced Video Coding (AVC). Although not shown in FIG. 1, in some aspects, the video encoder 20 and the video decoder 26 may be combined with an audio encoder and decoder, respectively, and include respective MUX-DEMUX or other hardware and software tools, for processing encoding of simultaneously audio and video data in a common data stream or in separate data streams. If possible, the MUX-DEMUX units can conform to the ITU H.223 multiplexer protocol or other protocols such as user datagram protocol (UDP).

The H.264 / MPEG-4 (AVC) standard was developed by the ITU-T Video Coding Expertise (VCEG) team with the ISO / IEC Moving Picture Expertise (MPEG) as a collaborative partnership product known as JVT . The H.264 standard is described in ITU-T Recommendation H.264, “Enhanced Video Coding for General Audiovisual Services,” issued by the ITU-T Research Group and dated March 2005, which may be referred to in this application as the H.264 standard, or technical H.264 standard, or H.264 / AVC standard or technical standard.

Joint Video Group (JVT) continues to work on expanding scalable video coding (SVC) to H.264 / MPEG-4 AVC. Technical standards for the development of the SVC extension are in the form of a joint project (JD). The Joint Scalable Video Data Model (JSVM) created by JVT implements tools for use in scalable video data that can be used in system 10 for various coding tasks described in this application. Detailed information on fine mesh coding with SNR scalability (FGS) can be found in documents for a joint project, for example, in joint project 6 (SVC JD6), Thomas Wiegand, Gary Sullivan, Julien Reichel, Heiko Schwartz and Mathias Wien, "Joint Draft 6 : Scalable Video Coding ", JVT-S 201, April 2006, Geneva, and Joint Project 9 (SVC JD9), Thomas Wiegand, Gary Sullivan, Julien Reichel, Heiko Schwartz and Mathias Wien," Joint Draft 9 of SVC Amendments, " JVT-V 201, January 2007, Marrakech, Morocco.

In some aspects for the case of broadcast video, the methods described herein can be applied to advanced H.264 video coding for the delivery of real-time video services in terrestrial mobile multimedia multi-subscriber (MT3) systems using the direct air interface technical standard only Forward Link Only Air Interface Specification for Terrestrial Mobile Multimedia Multicast (FLO), which will be published under the name TIA-1099 Technical Standard (FLO Technical Standard), for example, a server A wireless broadcast video or portable wireless device. The FLO technical standard includes examples defining the syntax, semantics, and bitstream decoding processes suitable for the FLO radio interface. Alternatively, video data may be broadcast in accordance with other standards, such as DVB-H (Digital Video Broadcasting - Portable), ISDB-T (Digital Broadcasting and Integrated Services - Terrestrial) or DMB (Digital Multimedia Broadcasting). Therefore, the source device 12 may be a mobile wireless terminal, a video streaming server, or a video broadcast server. However, the methods described in this application are not limited to any particular type of broadcast, multicast, or point-to-point transmission system. In the case of broadcast transmission, source device 12 may transmit multiple video channels to a plurality of receiver devices, each of which may be similar to receiver device 14 of FIG. 1.

Video encoder 20 and video decoder 26 can each be implemented as one or more microprocessors, digital signal processors (DSP), specialized integrated circuits (ASIC), user-programmable logic elements (FPGA), discrete logic, software, hardware, software hardware or any combination of the above. So, each video encoder 20 and video decoder 26 can be implemented, at least in part, in the form of a chip or integrated circuit (IC) device and is included in one or more encoders or decoders, any of which can be an integral part of the combined encoder / decoder (CODEC) in the corresponding mobile device, subscriber device, broadcast device, server, or the like. In addition, source device 12 and receiver device 14 may each include respective modulation, demodulation, frequency conversion, filtering, and amplification components for transmitting and receiving encoded video data, including radio frequency (RF) components and antennas sufficient to support wireless connection. However, to simplify the drawing, such components are not shown in FIG. 1.

A video sequence includes a sequence of video frames. For encoding video data, the video encoder 20 operates on blocks of pixels within individual video frames. Video blocks may have fixed or variable sizes and may vary in size according to a given coding standard. Each video frame includes a sequence of sections. Each section may include a sequence of macroblocks that can be arranged in sub-blocks. For example, the ITU-T H.264 standard supports intra prediction in blocks of various sizes, such as 16 by 16, 8 by 8, 4 by 4 for luminance components and 8 × 8 for color components, as well as inter prediction in blocks of various sizes, such as 16 by 16, 16 by 8, 8 by 16, 8 by 8, 8 by 4, 4 by 8 and 4 by 4 for luminance components and corresponding scaled sizes for color components.

Smaller blocks can provide better resolution and can be used in areas of the video frame that contain a higher level of detail. In general, macroblocks (MB) and various subblocks can be considered video blocks. In addition, a section can be considered a sequence of video blocks, such as MB and / or sub blocks. Each section may be an independently coded unit. After the prediction, a conversion can be performed on the 8 × 8 or 4 × 4 residual block, and an additional conversion can be applied to the 4 × 4 block DC coefficients for the color or luminance components if the intra prediction mode is used for 16 × blocks l6.

The video encoder 20 and / or video decoder 26 of the system 10 of FIG. 1 can be configured to employ variable length adaptive coding (VLC) techniques with memory efficiency and low complexity described in this application. In particular, the video encoder 20 and / or the video decoder 26 may include, respectively, a statistical code encoder and a statistical code decoder, which apply at least some of these methods to reduce memory usage, excessive processing, processing complexity, bandwidth occupancy, areas of data storage and / or power consumption.

FIG. 2 is a block diagram illustrating an example of the video encoder 20 of FIG. 1. The video encoder 20 may be made, at least in part, in the form of one or more integrated circuit devices, which collectively may be referred to as an integrated circuit device. In some aspects, video encoder 20 may be an integral part of a portable wireless communications device or broadcast server. Video encoder 20 may perform internal and external encoding of blocks within video frames. Internal coding relies on spatial prediction to reduce or eliminate spatial redundancy in the video data within a given video frame. External coding relies on temporal prediction to reduce or eliminate temporal redundancy in video data within adjacent frames of a video sequence. In the case of external coding, video encoder 20 performs motion estimation to track the movement of respective video blocks between adjacent frames.

As shown in FIG. 2, the video encoder 20 receives the current video block 30 within the video frame to be encoded. In the example of FIG. 2, the video encoder 20 includes a motion estimation module 32, a reference frame storage 34, a motion compensation module 36, a block transform module 38, a quantization module 40, a quantization module 42, an inverse transform module 44, and a statistical encoding module 46. Statistical encoding unit 46 may access one or more data structures stored in memory 45 to obtain data suitable for encoding. To filter blocks to remove blocking artifacts, a cyclic blocking reduction filter (not shown) can be used. The video encoder 20 also includes an adder 48 and an adder 50. Figure 2 shows the temporal prediction components of the video encoder 20 for external coding of video blocks. The video encoder 20 may also include spatial prediction components of the internal coding of some video blocks, although these components are not shown to simplify the drawing.

Motion estimation module 32 compares video block 30 with blocks in one or more adjacent video frames to generate one or more motion vectors. The adjacent frame or frames may be extracted from the reference frame storage 34, which may contain any type of memory or storage device for storing video blocks reconstructed from previously encoded blocks. Motion estimation can be performed for blocks of varying sizes, for example, 16 × 16, 16 × 8, 8 × 16, 8 × 8 or smaller blocks. Motion estimation module 32 identifies one or more blocks in adjacent frames that most closely match the current video block 30, for example, based on a cost-distortion model, and determines the offset between blocks in adjacent frames and the current video block. On this basis, motion estimation module 32 creates one or more motion vectors (MV) that indicate the magnitude and trajectory of the displacement between the current video block 30 and one or more matching blocks of reference frames used to encode the current video block 30.

Motion vectors can have an accuracy of half or a quarter of a pixel, and even higher accuracy, which allows the video encoder 20 to track the movement with higher accuracy than integer pixel locations, and to obtain a better prediction block. When motion vectors with fractional pixel values are used, interpolation operations are performed in the motion compensation unit 36. Motion estimation module 32 identifies the best blocking and motion vector or vectors for a video block using certain criteria, such as a cost-distortion model. For example, in the case of bidirectional prediction, you can get something more than a motion vector. Using the resulting blocking and motion vectors, the motion compensation module 36 forms a prediction video block.

The video encoder 20 forms a residual video block by subtracting the prediction video block obtained by the motion compensation module 36 from the original, current video block 30 in the adder 48. The block transform module 38 applies a transform, such as the integer 4 × 4 or 8 × 8 transform used in H. 264 / AVC, to the residual block with obtaining the coefficients of the residual converted block. Quantization module 40 quantizes the coefficients of the residual transformed block to further reduce the bit rate. The statistical coding unit 46 statistically encodes the quantized coefficients to further reduce the bit rate.

The statistical coding unit 46 acts as a variable length coding (VLC) block for applying VLC coding to the quantized block coefficients. In particular, statistical encoding module 46 may be configured to perform VLC coding of coefficients of digital video blocks using adaptive VLC techniques with memory efficiency and low complexity, as described herein. Thus, the various encoding processes described in this application can be performed in the statistical encoding unit 46 to perform encoding of the video data. Alternatively, such a statistical coding unit 46 may perform the processes described herein to encode any data, including, in particular, video data, image data, voice and audio data. In general, video decoder 26 performs inverse operations, including VLC decoding, for decoding and recovering encoded video data, as will be described later with reference to FIG. 3.

The inverse quantization unit 42 and the inverse transform unit 44 apply inverse quantization and inverse transform respectively to reconstruct the residual block. The adder 50 adds the reconstructed residual block to the motion compensated prediction block obtained by the motion compensation unit 36 to obtain a reconstructed video block for storing reference frames in the storage 34. The reconstructed video block is used by the motion estimation module 32 and the motion compensation module 36 to encode the block in the next video frame.

3 is a block diagram illustrating an example of a video decoder 26. The video decoder 26 may be made, at least in part, in the form of one or more integrated circuit devices, which collectively may be called an integrated circuit device. In some aspects, video decoder 26 may be an integral part of a portable wireless communications device. Video decoder 26 may perform internal and external decoding within video frames. As shown in FIG. 3, video decoder 26 receives an encoded video bitstream that has been encoded by video encoder 20. In the example of FIG. 3, video decoder 26 includes a statistical decoding unit 52, a motion compensation unit 54, an inverse quantization unit 56, an inverse transform unit 58, and a reference frame storage 62. The statistical decoding unit 52 may access one or more data structures stored in the memory 51 to obtain the data required for encoding. Video decoder 26 may also include a cyclic deblocking filter (not shown) that filters the output from adder 64. Video decoder 26 also includes adder 64. FIG. 3 shows temporal prediction components of video decoder 26 for external decoding of video blocks . Video decoder 26 may also include spatial prediction components for internally decoding some video blocks, although these components are not shown in FIG. 3.

The statistical decoding unit 52 receives the encoded video bitstream and decodes the quantized residual coefficients, macroblock coding mode, and motion information from the bitstream, which may include motion vectors and blocking. Thus, the statistical decoding unit 52 performs the function of a VLC decoding unit. For example, to decode the quantized residual coefficients from the encoded bitstream, the statistical decoding module 52 of FIG. 3, like the statistical encoding module 46 of FIG. 2, can perform adaptive VLC decoding of coefficients of blocks of digital video data, as described herein. However, entropy decoding unit 52 may perform VLC decoding inversely with respect to entropy encoding unit 46 in FIG. 2 to extract quantized block coefficients from the encoded bitstream. Thus, the various decoding processes described herein can be implemented in a statistical decoding unit 52 for decoding video data. Alternatively, such a statistical decoding module 52 may perform the processes described herein to decode any of a variety of data types, including but not limited to video data, image data, voice and audio data. In any case, the variable-length encoding result performed by the statistical decoding unit 52 may be provided to the user, stored in memory, and / or transferred to another device or processing unit.

The motion compensation module 54 receives the motion and blocking vectors and one or more reconstructed reference frames from the reference frame storage 62 to obtain a video prediction block. The inverse quantization module 56 performs inverse quantization, i.e., dequantization of the quantized block coefficients. The inverse transform unit 58 applies an inverse transform to these coefficients, for example, an inverse DCT or an inverse 4x4 or 8x4 integer transform to produce residual blocks. The video prediction blocks are then summed by the adder 64 with the residual blocks to form the decoded blocks. A deblocking filter (not shown) may be used to filter decoded blocks to remove blocking artifacts. Then the filtered blocks are placed in the storage 62 reference frames, which provides reference frames for decoding subsequent video frames and also creates decoded video data to control the display device 28 (Fig. 1)

Memory efficient coding of variable length codes

In accordance with a first general aspect of the invention, a method for efficiently utilizing memory for variable length coding to support compact data structures is described in detail below. This method should not necessarily rely on any particular code construction scheme, such as Huffman, Shannon, Shannon-Fano, Hilbert-Moore, or other codes. This method assumes, however, that the code is constructed for a source with monotonically increasing character probabilities. More specifically, it is assumed that codewords have monotonically decreasing (or at least non-increasing) lengths and that codewords of the same length have the same lexicographic order as the characters in the input alphabet that they represent.

This method, when applied to encoding video data or in other applications, uses the above properties to uniquely describe such code with a very compact data structure. As described above, the data structure may contain, for each valid codeword length, that is, for a level with external nodes in the code tree, the following information:

a) the particular value of the lexicographically smallest (or largest) codeword at the current level of the code tree,

b) the number of bits in a partial value,

c) the meaning of the character corresponding to the lexicographically smallest (or largest) codeword, and

d) a metric that tells the decoder to skip a certain number of bits before moving on to the next level of the code tree.

Encoding and decoding processes can use this structure to identify the level of the code tree corresponding to the code word that you want to receive (or decode), and then directly calculate the value of the code word (or decoded symbol).

This method may allow the use of significantly less memory for encoding and decoding due to the fact that only partial values of the lexicographically smallest code words are stored. A data structure may refer to any of a variety of known data structures, such as tables, linked lists, binary trees, basis trees, structureless files, or the like, and may be stored in any of a variety of different storage devices, such as random access memory (RAM), read-only memory (ROM), or both at the same time. The data structure may be stored in such a memory inside the encoder or decoder, for example, in the memory 45 or in the memory 51 shown in FIGS. 2 and 3, respectively, in which case at least some of the levels of the coding tree include codewords located in lexicographic order relative to the order of the values of the characters represented by the code words. Accordingly, each of the base codewords is the lexicographically smallest codeword at the corresponding level of the coding tree. In addition, this method allows the use of phased access to data in the bitstream, as a result of which the buffer of the bitstream can be represented by a reasonably short register. For example, a 32-bit register may be sufficient even for very long codes. The register can be updated at conveniently selected intervals (for example, after 8 bits), which further reduces the implementation complexity. In general, this method, in its various aspects, can significantly reduce the complexity of the representation and encoding / decoding of variable-length codes, which may allow developers of compression algorithms to use large, more efficient codebooks.

An overview of variable length codes will now be given in order to help illustrate the methods described in this application. Variable-length codes are a fundamental tool in data compression. In general, variable-length codes are used to represent a sequence of characters with some well-known and usually extremely unbalanced distribution. Such sequences may be represented by binary sequences, or codes of significantly shorter length. Reducing the length is provided by replacing the more common characters with shorter codes, and less private characters with longer codes.

Examples of variable length codes used to compress data are Huffman codes, such as those described in D.A. Huffman. A method of the construction of mminimum-redundancy codes. Proc. IRE, vol. 40, pp. 1098-1101. Sept. 1952, Shannon codes, for example, described in C.E. Shannon, A mathematical theory of communication, Bell Syst. Tech J. Vol. 27. pp. 379-423, July 1948, Shannon-Fano codes, for example, described in M R. Fano, The transmission of information, Res. Lab. Electronics, Massachusetts Inst. of Technology, Cambridge, Mass. Tech, Rep. No. 65, 1949, and Hilbert-Moore codes, for example, described by E. N. Gilbert and E. F. Moore, Variable-Length Binary Encoding, Bell. Syst. Tech. J., Vol. 7, pp. 932-967, 1959 (sometimes also called Shannon-Fano-Elias codes).

The codes described above belong to the class of non-prefix codes, for example, as described by T. Cover and J. Thomas, Element of Information Theory, Wiley, 1991. Fig. 4 is a diagram illustrating an example of a binary tree. The codes described above are conveniently represented in the form of a binary tree, such as that shown in FIG. 4. Thus, the encoder can encode symbol values in accordance with the encoding tree. Values according to the tree can represent any of a variety of data, such as video data, image data, voice or audio data. Each internal node in such a tree corresponds to a binary symbol whose value equal to 0 causes a step to the right, and a value equal to 1 causes a step to the left to a child node in the tree. The topmost node is called the root node, and it is the node from which encoding / decoding begins.

Each external node in the tree is a node where the encoding / decoding process is resumed, resulting in either a word, that is, a sequence of bits from the root node to the current node, or a decoded symbol value associated with the current code word. In the example coding tree of FIG. 4, there are sixteen codewords corresponding to characters indexed from 0 to 15. In this example, the shortest codeword is 1 bit long and the longest codewords are 10 bit long. The number of levels containing external nodes (code words) in this tree is 7, namely, 1st, 3rd, 4th, 6th, 7th, 9th and 10th levels.

As shown in Fig. 4 below, let n denote the number of external nodes in the encoding tree (and, accordingly, the number of codewords in the code), let L denote the length of the longest codeword, and let K denote the number of levels filled with external nodes in the tree coding. In further consideration, the designation "O-large" is used, introduced by P. Bachmann. For example, the expression y (n) = O (x (n)) means the presence of some constant M> 0, so that y () ≤M | x (n) | for all sufficiently large n.

To support the encoding process, an encoder and a decoder, such as a statistical encoding module 46 and a statistical decoding module 52, generally should store a binary tree in a computer memory, such as a memory 45 or a memory 51. In addition, the encoding and decoding processes should include a bitwise (i.e., node-wise) traversal of the encoding tree stored in memory. Therefore, for such an implementation, storage costs must be O (n), and the number of steps can be up to O (L). However, in some special cases where the encoding trees have some special structure, there may be more efficient ways to store such code structures and perform encoding or decoding operations.

For example, looking at the code presented in the example coding tree of FIG. 4, it can be noted that the codeword length does not decrease and that all codewords at the same level of the coding tree have adjacent values. For example, the codewords at the 4th level of the tree in FIG. 4 are longer than the codewords at the 3rd level of the tree, namely 0001 compared to 011, 010, 001 and 000. In addition, the codewords at 3- m level have adjacent values 011, 010, 011, 000. Therefore, instead of storing all the codes, it may be sufficient to store only the smallest or largest codeword for each level of the coding tree, that is, as the base codeword from which it can be calculated all adjacent codewords.

The above observation is the key to understanding how to decode variable-length codes based on their conversion to the so-called canonical form, for example, as described in A. Moffat and A. Turpin, On the Implementation of Minimum-Redundancy Prefix Codes, IEEE Trans. Communications, 45 (10) (1997) 1200-1207. In simple words, the canonical code has a non-decreasing distribution of lengths and supports the lexicographic order with respect to the indices assigned to its nodes. It is quite simple to show that any given source can be reordered in such a way that the resulting code will have the above properties.

For example, the code shown in the coding tree of FIG. 4 is a reordered code for a non-monotonous source, as shown in Table 1 below. In particular, Table 1 is an example of a canonical variable-length code that has been reordered.

Table 1
Variable Length Canonical Code Example Symbol Probability Character index after reordering Code length The code 0 0.6561 fifteen one one one 0,0729 12 3 011 2 0,0729 13 3 010 3 0.0081 5 7 0000101 four 0,0729 fourteen 3 001 5 0.0081 6 7 0000100 6 0.0081 10 6 000011 7 0,0009 one 10 0000000001 8 0,0729 eleven four 0001 9 0.0081 7 7 0000011 10 0.0081 8 7 0000010 eleven 0,0009 2 9 000000011 12 0.0081 9 7 0000001 13 0,0009 3 9 000000010 fourteen 0,0009 four 9 000000001 fifteen 0.0001 0 10 0000000000

In table 1 above, the symbol 0 has the highest probability, followed by the symbols 1 and 2. However, the symbol 3 has a lower probability than 4, and 4 and 8 have the same probability as 1 and 2. After reordering, all the probabilities characters monotonically increase (do not decrease) and lead to the canonical code represented by the encoding tree in figure 4. The so-called Moffat-Turpin algorithm described by A. Moffat and A. Turpin, On the Implementation of Minimum-Redundancy Prefix Codes, IEEE Trans. Communications, 45 (10) (1997) 1200-1207, provides a method for decoding canonical codes. The methods described in this application may provide an improvement to the Moffat-Turpin algorithm. Other algorithms, such as those described by J. B. Connell, A Huffman-Shannon-Fano Code, Proc. IEEE, July 1973, 1046-1047, and A. Brodnik and S. Carlsson, Sub linear Decoding of Huffman Codes Almost in Place, DIMACS Workshop on Codes and Trees: Algorithmic and information Theoretic Approaches, October 1998, Rutgers University, DIMACS Center, NJs are similar to the Moffat-Turpin algorithm, and can also be improved by using the disclosed methods in a similar manner.

The Moffat-Turpin algorithm for decoding variable-length codes is described below. Suppose that the input alphabet A contains n letters A = {α ₀ , ... α _n-1 }, the reordering i: A → {0, ..., n-1} is applied, so that the resulting probabilities satisfy the condition: p ₀ ≤p ₁ ≤ ... ≤p _n-1 . Then, the Huffman algorithm or another algorithm for constructing the minimum redundancy, which assigns the length of codewords l _{i to} each index 1≤i≤L, where L is the length of the longest codeword, can be applied. The result is "fill numbers" m ₁ , that is, the number of code words of length l.

Using these parameters, the so-called "base" values are calculated for each level in the tree according to the following formula:

, (1≤l≤L)

, (1≤l≤L)

These basic codeword values represent the lexicographically smallest canonical codewords at each level of the tree. Knowing the base value of the codeword base [l], we can now calculate the j + 1th codeword among m _l codewords of length l:

base [l] + j.

For the decoder to work, it is more convenient to store a left-aligned version of the base value of the code word:

lj_base [l] = base [l] 2 ^Wl ,

where W is the length of the bit buffer or register used to store the most recently loaded bits from the bitstream. It is assumed that W≥L.

Finally, in addition to the basic codeword values, the Moffat-Turpin decoder also stores the indices of the corresponding symbols in a reordered array. These indices are stored as offset values [l] for each filled level in the tree. An example of the complete structure supported by the Moffat-Turpin algorithm for the code represented by the tree in FIG. 4 is shown in Table 2 below.

table 2
The structure of the Moffat-Turpin decoder for the code in figure 4 i Lj_base [i] (W = 16) Level [i] Offset [i] 0 1 000000000000000 one fifteen one 001 0000000000000 3 12 2 0001 000000000000 four eleven

3 000011 0000000000 6 10 four 0000001 000000000 7 5 5 000000001 0000000 9 2 6 0000000000000000 10 0

An example of pseudo-code for the implementation of the Moffat-Turpin decoding algorithm using the structure from table 2 is presented below in table 3.

Table 3
Moffat-Turpin Decoding Algorithm Line Team Comment one V = bitstream buffer get the last W bits from the bitstream 2 For (i = 0; i <K; i ++) { 3 if (lj_base [i] <= V) search level four break; containing the current codeword 5 } 6 l = level [i]; get the length 7 scroll_bitsream (l); scroll bit stream by l bits 8 Symbol = offset [i] +
((V-base [i]) >>(Wl)); decode character

From Table 3 above, it can be seen that the entire decoding process includes up to K (W-bit) comparisons of the current bitstream buffer with left-aligned values of the base codewords, followed by a simple direct calculation of the decoded symbol index. You can also see that the lj_base [] array used by such a structure requires O (K · W) bits of memory, which could be a problem if the codeword were long, since W should be used so that W≥1.

In the example from Table 3, the decoder receives W bits from the bitstream and as V and compares V with the base codewords (lj_base [i]) for subsequent levels i of the encoding tree. When it is determined that the base codeword is less than or equal to V, the search for the codeword level is terminated. The decoder then determines the length associated with level i, scrolls the bitstream by l bits, and decodes the symbol. In particular, the decoded symbol is determined by the sum of the offset value for level i and the difference between the code word V from the data stream and the base code word for level i shifted to the right by W-I bits.

In the general case of implementing the Moffat-Turpin decoding, the inverse mapping is: i ^-1 : {0, ..., n-1} → A. In this case, the last operation becomes the most memory-intensive, since it requires an area of size O (n). However, in many cases encountered in practice, such as situations involving the sequential coding or output of transformations or prognostic parameters, the sources to be encoded are already ordered. Consequently, the memory used by the lj_base [] array in the Moffat-Turpin structure becomes a major factor in overall storage costs.

In accordance with a first general aspect, the methods disclosed herein provide improvements that allow even more compression of data structures used in the Moffat-Turpin algorithm or other algorithms and support phased coding of variable length codes. Now consider these improvements in more detail. From table 2 it clearly follows that the values of lj_base [l] have a large number of initial bits, when viewed from right to left. Therefore, partial values of the base codewords are the removal of a fixed number of leading bits from the base codewords. In most cases, the deleted start bits are zeros. This number of zeros increases monotonically as the coding tree expands to deeper levels. Thus, if the bits are sequentially decoded, starting from the very first level of the encoding tree and moving down, you can skip several initial zero bits without any effect on the decoding. Skipping at least a few leading zeros in the methods described in this application allows us to obtain highly compressed data structures and stage-by-stage decoding of codes of variable length.

It is possible, however, that after deleting the initial bits, some legitimate codes at the lower levels of the encoding tree may enter the region of the deleted initial bits. Accordingly, in order to prevent the loss of such codes, a table of pass indicators is provided. The skip rate tells the decoder to skip a few bits in the bitstream, which will be decoded before moving to the selected level of the encoding tree. In particular, the skip rate may instruct the decoder to skip a fixed number of bits, for example, 8 bits, in the bitstream of the codewords before moving to the selected level of the encoding tree. Thus, the partial value of the base codeword at the selected tree level is based on the partial offset of the base codeword by a fixed number of bits. Without this shift, the base codeword at the selected tree level will extend, at least in part, to the remote number of start bits.

Table 4 below illustrates an example implementation of an encoding process in which leading zeros are removed, in accordance with one aspect of the present invention, to further compress the data structure used to represent and process code words. In the example of Table 4, leading zeros are deleted in increments of 8, which is convenient for controlling the bitstream buffer in regular 8-bit / byte computers. To control the removal of leading zeros, an additional indicator table (skip_8 [i]) is provided, as indicated above. Thus, table 4 as a whole corresponds to table 2, but the leading zeros from each codeword are removed and a column with a skip indicator is added.

Table 4
Modified Moffat-Turpin Decoding Structure i r_lj_base [i] (W = 8) skip_8 [i] r_level [i] offset [i] 0 1 000 000 000 0 one fifteen one 001 00000 0 3 12 2 0001 0000 0 four eleven 3 000011 00 0 6 10 four 0000001 0 one 7 5 5 1 000 000 000 0 9-8 = 1 2 6 00 000 000 0 10-8 = 2 0

In the example of Table 4, the value r_lj_base (i) represents the value of the base codeword at each index position, the value r_level [i] indicates the level inside the coding tree for the index position and the length of codewords at this level, the offset [i] value indicates the number of leading zeros to the right to the left for the value of the base codeword, and the value skip_8 [i] indicates whether the decoder should skip 8 bits to obtain the next value of the base codeword, with 1 indicating a gap and 0 indicating no gap. This skip operation periodically updates the bit buffer at selected intervals, which allows the decoder to identify codewords that would otherwise be lost after the removal of leading zeros. For example, if the rightmost eight leading zeros of a left-aligned codeword are deleted, the codeword that extends to the rightmost eight bits would be partially or completely lost. Accordingly, skipping the left-most eight bits in response to the skip rate will move the codeword to the region of undeleted bits, thereby preserving the codeword for decoding.

Therefore, the skip indicator signals when the decoder should skip ahead by the specified skip step for the next code level, that is, 8 in this example of table 4. For example, in table 2, the values of the basic code words are positions 5 and 6 (tree levels 9 and 10) are equal, respectively, 0000000010000000 and 0000000000000000, respectively. After removing the rightmost eight leading zeros (left justified) from these values of the base codewords, it is necessary that the decoder jumps forward eight bits so that the actual value of the base codeword (0000000010000000) is not lost when removing the eight leading zeros. Instead, the actual value of the base codeword (0000000010000000) is converted to another value of the base codeword (10000000) by skipping the first eight bits (00000000) and then deleting the rightmost eight leading zeros.

Due to the removal of leading zeros, the width of the modified lj_base [i] array is much smaller. For example, in the code in table 4, the width W of the modified array lj_base [i] W = 8, instead of W = 16 in the case of table 2. An example implementation of an algorithm that uses such an additional skip table to periodically update the bit buffer is shown in Table 5 below. The algorithm, constructed as shown in table 5, can be performed with the ability to support very long codewords and very compact value tables (lj_base) of the base codewords.

Table 5
Modified Moffat-Turpin Decoder Algorithm Line Team Comment one V = bitstream_buffer get the last W bits from the bitstream 2 for (i = 0; i <K; i ++) { 3 if (lj_base [i] <= V) level search four break; containing the current codeword 5 if (skip_B [i]) Should I skip the next B bits 6 V = scroll_bitsream (B); scroll bitstream by B bits 7 } 8 l = level [i] get residual length 9 scroll_bitstream (l) scroll bit stream by l bits 10 Symbol = offset [i] + ((V-base [i]) >>
(Wl)); decode character

As shown in table 5, the decoder receives the last W bits from the bit stream represented by the value V = bitstream_buffer. The decoder then searches for coding tree levels i for the value lj_base [i] of the base codeword that is less than or equal to codeword V from the bitstream buffer. If the current level i of the tree corresponds to the skip level (skip_B [i]), for example, as shown in table 5, then the decoder scrolls the bit stream to the right by B bits, for example, by 8 bits in some embodiment, so that the codeword is at the next level , which the decoder is looking for, could be saved rather than lost when deleting B of the rightmost start bits.

After determining the residual length l = level [i] for codewords at the current tree level, for example, as shown in table 5, the decoder scrolls the bitstream by length l. The decoder then directly calculates the symbol index based on the offset amount for the current level i and the difference between the contents V of the data stream buffer and the base codeword for the current level i shifted to the right by W-I bits.

The decoder decodes the codeword from the bitstream of the codeword using a stored data structure defining partial values for the base codewords, skip rate, values represented by the base codeword, and lengths (namely, the number of bits) of the partial values in the base codewords. In general, a processor associated with a decoder, such as statistical decoding unit 52, searches for encoding tree levels of a selected one of the partial values of the base codewords that is less than or equal to the codeword from the codeword bitstream. The processor skips several bits in the bitstream of the codeword before moving to the selected level of the coding tree in response to the skip rate and calculates one of a plurality of values corresponding to the codeword based on the difference between the selected from the partial values of the base codewords that is less than or equal to the codeword , and a codeword and an index of a selected value from partial values of the base codewords that are less than or equal to the codeword. The processor generates a decoding result for storage, transmission to another device or processing module, or presentation to the user. For example, the decoded results can be used to control the display device 28 to represent video or a still image and / or the audio output device to produce sound or speech.

In the example in Table 5, the decoder incrementally updates the bitstream buffer through a skip operation to save codewords that would otherwise be lost after the removal of leading zeros. In addition, the values of the basic codewords that the decoder compares at each code level can be significantly shorter. The following describes the potential scope for reducing the length of the meaning of the base codewords. To analyze the dynamic range of such quantities in the modified algorithm described in this application, the difference between two adjacent levels is determined from the following formula:

If i is the index of the next non-empty level, then:

lj_base [l] - lj_base [l + i] = 2 ^Wl + m _{l + i} 2 ^{W- (l + i)} .

Here, the main quantity of interest is: m _{l + i} 2 ^-i , which affects this difference. In the simplest case, when i = 1, it is clear that this difference simply depends on the number of external nodes, and, therefore, W can be chosen in such a way that:

W ≥ max log ₂ (m _l ),

and this value in cases of interest in practice is much less than L. This difference should be especially large for extremely unbalanced distributions.

кодовых слов, а это значит, что требуется использовать приблизительно H(p)m битов, чтобы различить кодовые слова, где H(p) - функция энтропии, например, описанная в работе T. Cover and J. Thomas, Elements of Information Theory, Wiley, 1991.For example, if the input symbols are blocks of m bits with Bernoulli probabilities p ^m (1-p) ^mk , then the most populated level should contain approximately

code words, which means that you need to use approximately H (p) m bits to distinguish code words, where H (p) is the entropy function, for example, described in T. Cover and J. Thomas, Elements of Information Theory, Wiley, 1991.

битов, причем известно, что для асимметричных распределенийOn the other hand, the longest codeword in this case will have approximately

bits, and it is known that for asymmetric distributions

,

,

- специальный случай энтропии Реньи, например, описанный в работе W. Szpankowski, Average Case Analysis of Algorithms on Sequences. (New York, John Wiley & Sons, 2001). Эта разность может быть произвольно большой при p → 0 или p → 1.Where

- a special case of Renyi entropy, for example, described in W. Szpankowski, Average Case Analysis of Algorithms on Sequences. (New York, John Wiley & Sons, 2001). This difference can be arbitrarily large as p → 0 or p → 1.

From the above consideration it follows that the proposed method should be effective in processing large, asymmetric code structures. It is traditionally difficult to work with such structures using traditional / existing methods, and in many cases, engineers resort to simplifications that affect the compression efficiency of codes to make them more practical.

For example, the very popular Golomb codes, described, for example, in S. Golomb, "Run-length coding," IEEE Trans. Inform. Theory, vol. IT-12, pp. 399-401, July 1966, and R. Gallager and D. van Voorhis, "Optimal source codes for geometrically distributed integer alphabets," IEEE Trans. Inform. Theory, vol. IT-21, pp. 228-230, Mar. 1975, are variable-length codes with a particularly simple structure, but they are optimal only for one class of geometric distributions and only for a countable number of parameter values of such distributions. Engineers strive to use them even for significantly different distributions, motivating this mainly with considerations of complexity. Such solutions may become at the same time not entirely optimal and difficult to expand or change due to the performance limitations inherent in such codes.

Another solution related to the construction of Lynch-Davisson codes described in T. J. Lynch, Sequence time coding for data compression, Proc. IEEE (Lett.), 54 (1966) 1490-1491, and L. D. Davisson, Comments on Sequence time coding for data compression, Proc. IEEE (Lett.), 54 (1966) 2010-2011, is to split the codes into two parts, with only the first being subjected to variable length encoding, and the other being transmitted as an extension using a fixed number of bits. Unfortunately, with such splitting, efficiency is lost, and the loss sometimes reaches 1.5-2 bits per character.

A more sophisticated version of the codebook splitting method was developed, called alphabetical grouping, described, for example, in the works of Boris Ryabko, Jaakko Astola, Karen Egiazarian, Fast Codes for Large Alphabets, Communications in Information and Systems, v.3, n. 2, pp. 139-152, and Boris Ryabko, Jorma Rissanen, Fast Adaptive Arithmetic Code for Large Alphabet Sources with Asymmetrical Distributions, IEEE Communications Letters, v. 7, no. 1, 2003, pp. 33-35. However, this approach is also implemented due to some loss in compression efficiency.

In contrast to the above methods, the methods described in this application can be performed with the possibility of fully preserving the structure and optimality of the code and therefore can be useful for a wide range of practical applications in the field of data compression and encoding, for example, when encoding and decoding digital video data, data images, sound and voice data.

Binary Adaptive Block Coding

The following describes in detail an example of a low-complexity adaptive variable-length coding method for binary sequences generated by memoryless sources in accordance with a second general aspect of the present invention. This disclosed method may implement universal block codes constructed for a set of contexts identified by the number of non-zero bits in previous bits in a sequence. This second general aspect of the invention can be implemented or provided independently or in combination with the first general aspect described above with respect to the generation of very compact data structures. Data structures can refer to a wide range of data structures, such as tables, linked lists, binary trees, basis trees, structureless files, or the like, and can be stored in any of a variety of different storage devices, such as many kinds of random access memory (RAM). , read only memory (ROM), or both. The data structure may be stored in such memory within the encoder or decoder. In accordance with this second general aspect, a method for adaptive encoding and decoding having a reduced degree of complexity can rely, at least in part, on a formula for the asymptotic redundancy of such codes, which refines the estimate described by RE Krichevsky and VK Trofimov, The Performance of Universal Encoding, IEEE Trans. Information Theory, 27 (1981) 199-207.

Data compression algorithms convert input bit sequences having some unknown distribution into a decoded bit stream. Data compression is used, for example, in the development of encoders of image data and video data, in scalable (based on sectioning) coding of the spectrum in audio encoders and in other applications. In most of these cases, the bits to be encoded are taken from the values created by various signal processing tools, such as transforms, prediction filters, etc., which means that they are already decorrelated and that there is no memory assumption for such a source is warranted.

The most commonly used options for implementing such binary adaptive algorithms are usually based on arithmetic codes using some approximations and abbreviations to reduce their complexity. Two well-known examples of such algorithms are the Q-encoder algorithm described by WB Pennebaker, JL Mitchell, GG Langdon, Jr., RB Arps, An overview of the basic principles of the Q-Coder adaptive binary arithmetic coder, IBM J. Res . Dev., 32 (6) (1988) 717, which is used in the JBIG image encoding standard, or the CABAC algorithm described in D. Marpe, H. Schwartz, and T. Wiegand, Context-Based Adaptive Binary Arithmetic Coding in the H. 264 / AVC video compression standard, IEEE Trans, on CSVT, 13 (7): 620 636, July 2003, which is used in the ITU-T H.264 / MPEG AVC video compression standards.

In accordance with this second general aspect of the invention, an alternative embodiment uses an array of Huffman codes for multiple estimated densities and indexed by the number of non-zero bits in previous blocks (contexts) in the sequence. From the point of view of both efficiency and implementation, such a method can provide the required compression characteristics when using even blocks of bits of moderate sizes, for example, n = 8 ... 16, and when using appropriate memory volumes, for example, from 1.5 kilobytes to 5 kilobytes .

This method can be considered in the context of a memoryless source that creates characters from the binary alphabet {0,1} with probabilities p and q = 1-p, respectively. If w is a word with a length n created by this source, then its probability:

Pr (w) = p ^k q ^nk , (1)

where k denotes the number of units in this word. The value of k may also be called the weight of w.

The block code ϕ is an injective mapping between words w of length | w | = n and binary sequences (or code words) ϕ (w):

ϕ: {0,1} ⁿ → {0,1} ^* ,

where the code words ϕ (w) are a uniquely decoded set, as described, for example, in T. M. Cover and J. M. Thomas, Elements of Information Theory, (John Wiley & Sons, New York, 1991).

As a rule, when the source (i.e., its probability p) is known, such a code is intended to minimize its average length or, in relative terms, its average redundancy:

.

.

In the above formula, H (p) = - plogp - qlogq denotes the source entropy.

Classical examples of codes and algorithms proposed to solve this problem include Huffman, Shannon, Shannon-Fano, and Gilbert-Moore codes and their variants. The characteristics of such codes are well studied; many methods of implementing such codes that are useful for practical application are known from the literature.

When the source is unknown, the best available option is to construct a universal ϕ ^* code that minimizes worst case redundancy for a certain class of sources, as described, for example, in BM Fitingof, Optimal Coding in the Case of Unknown and Changing Message Statistics, Probl. Inform. Transm., 2, (2) (1965) 3 {11 (in Russian) 1-7 (English Transl.), LD Davisson, Universal Noiseless Coding, IEEE Trans. Inform. Theory, 19 (6) (1973), 783-795, and RE Krichevsky and VK Trofimov, The Performance of Universal Encoding, IEEE Trans. Information Theory, 27 (1981) 199-207, and as follows:

.

.

An example of such a code can be constructed using the following word probability estimates:

,

,

where Γ (x) is a Γ-function, k is the weight of the word w, and n is its length. The above formula is described by R. E. Krichevsky and V. K. Trofimov, The Performance of Universal Encoding, IEEE Trans. Information Theory, 27 (1981) 199-207, and provides uniform (in ρ) convergence to true probabilities as n approaches infinity (n → ∞).

In a situation where the exact value of the source parameter is unknown, you can refer to the sequence of characters u created by the same source in the past. Such a sequence may be called a sample, and its length may be assumed to be equal to | u | = t bits. The challenge here is to construct a set of ϕ ^* _u codes indexed by different values of this sample so that their resulting average worst case redundancy is minimal, as shown below:

.

.

Such codes are called sample-based, or adaptive, universal block codes. This application describes private embodiments of adaptive block codes using the following probability estimates for words w of a given sample u:

(1)

(one)

where s is the sample weight u and t is its length.

The concept and analysis of sample-based codes using the rating function (1) above is described in R. E. Krichevsky in R. E. Krichevsky, Universal Data Compression and Retrieval. (Kluwer, Norwell, MA, 1993). Asymptotically, the average degree of redundancy of the adaptive block code is:

, (2)

, (2)

where n is the block size and t is the size of the samples.

From the above formula (2), it is obvious that when using samples of length t = O (n), the degree of redundancy of such codes can be reduced to O (1 / n), which in order of magnitude corresponds to the degree of redundancy of block codes of known sources. However, to fully understand the potential of such codes, it is necessary to know a more accurate expression for their redundancy, including terms that are affected by the choice of the actual algorithm for constructing the code, such as the Huffman, Shannon code or the like.

In accordance with this second general aspect, the present invention provides the following refinement of Krichevsky’s Theorem 2. In particular, Theorem 1 below refines the theorem on the average degree of redundancy for the adaptive block code ϕ ^* _u as follows:

Theorem 1: The average degree of redundancy of the adaptive block code ϕ ^* _u has the following asymptotic behavior (n, t → ∞):

(3A)

where n is the block size, and t is the sample size, p, q = 1-p are the probabilities of the symbols of the input source, and where:

the average redundancy of the code ϕ ^* _u relative to the estimated distribution in formula (1).

зависит от алгоритма. Однако для большого класса способов обеспечения минимальной избыточности, который включает в себя традиционные коды Хаффмана и Шеннона, величина этого члена ограничена следующим условием:Exact value behavior

Depends on the algorithm. However, for a large class of methods for ensuring minimal redundancy, which includes the traditional Huffman and Shannon codes, the value of this term is limited by the following condition:

,

,

and is characterized by oscillatory behavior, which may or may not converge to a certain constant value depending on the value of the parameter p. In addition, for small values of t and n, the following member can influence the redundancy of such codes:

which is a function of the source parameter p. 5 is a graph illustrating the degree of redundancy of an adaptive block code with asymptotic behavior. For short blocks / samples, the characteristics of such codes become sensitive to source asymmetry. A proof of this theorem can be found, for example, in Asymptotic average redundancy of adaptive block codes, Y. A. Reznik, W. Szpankowski, Proceedings of IEEE International Symposium on Information Theory (ISIT), 2003.

The following are examples of efficient algorithms for implementing the codes described above. In a memoryless model, the probability of the word w (or its estimate) depends only on its weight k, and not on the actual combination of its bits. Therefore, considering the set of all possible words of n bits, we can divide this set into n + 1 groups:

,

,

. Далее для удобства предполагается, что каждая группа W_n,k хранит слова в лексикографическом порядке. Значение I_n,k(w) обозначает индекс (позицию) слова w в группе W_n,k. Приведенная ниже таблица 6 служит примером кода, построенного для 4-битовых блоков с Бернуллиевыми вероятностями: p^kq^n-k, p=0,9.containing words with the same weight (k = 0, ..., n) and the same probability. The sizes of such groups are equal

. Further, for convenience, it is assumed that each group W _{n, k} stores words in lexicographical order. The value of I _{n, k} (w) denotes the index (position) of the word w in the group W _{n, k} . Table 6 below is an example of code built for 4-bit blocks with Bernoulli probabilities: p ^k q ^nk , p = 0.9.

Table 6
Sample code built for 4-bit blocks
with Bernoulli probabilities p ^k q ^nk , p = 0.9 Block w k I _{n, k} (w) Pr (w) Length Code ϕ (w) Subgroup 0000
0001
0010
0011
0100
0101
0110
0111
1000
1001
1010
1011
1100
1101
1110
1111 0
one
one
2
one
2
2
3
one
2
2
3
2
3
3
four 0
0
one
0
2
one
2
0
3
3
four
one
5
2
3
0 0.6561
0,0729
0,0729
0.0081
0,0729
0.0081
0.0081
0,0009
0,0729
0.0081
0.0081
0,0009
0.0081
0,0009
0,0009
0.0001 one
3
3
6
3
7
7
9
four
7
7
9
7
9
10
10 one
001
010
000011
011
0000001
0000010
000000001
0001
0000011
000100
000000010
0000101
000000011
0000000001
0000000000 0
one
one
3
one
four
four
5
2
four
four
5
four
5
6
7

The code example in Table 6 will be used to describe the structure of the proposed correspondence between the words in the group W _{n, k} and their code words in accordance with some aspects of the present invention. This code was built using a modification of the Huffman algorithm, in which additional steps were performed to ensure that codewords located at the same level have the same lexicographic order as the input blocks that they represent. It is known that such reordering is possible without loss of compression efficiency. Examples of previously known algorithms that used this idea of reordering include Huffman-Shannon-Fano codes and canonical codes described by Moffat and Turpin.

Fig. 6 is a diagram showing a coding tree that displays the structure of the example block code shown in Table 6. As expected, each group W _{n, k} consists of a maximum of two subgroups containing code words of the same length. In general, the code structure represented by the coding tree of FIG. 6 and the block codes from Table 6 defines the codeword groups and the first and second subgroups of codewords within each of the groups. Each of the groups includes code words representing values with the same weights. The first subgroup includes codewords having a first length, and the second subgroup includes codewords having a second length different from the first length. The codewords in each of the groups are ordered in lexicographic order relative to the values represented by the codewords, to facilitate encoding and decoding by direct calculation.

An example of a group is indicated in FIG. 6 by reference numeral 66. Examples of the first and second subgroups are indicated in FIG. 6 by reference numbers 68A, 68B. Similar groups and subgroups are provided for each weight within the coding tree. The group contains blocks having the same weight k. The subgroup contains blocks that have the same weight and the same level in the encoding tree. This follows from the fact that all words in the group W _{n, k} have the same probability and have the so-called fraternal property of Huffman codes. This observation holds true for Shannon codes, generalized Shannon codes, and, possibly, for other algorithms. As noted above, the group W _{n, k} includes a maximum of two subgroups containing code words of the same length, and can be represented as:

,

,

- наименьшая длина кода, которая может быть присвоена блокам из группы W_n,k. Кроме того, поскольку слова в группе W_n,k следуют в лексикографическом порядке, то разделение на W_n,k,l и W_n,k,l+1 - это просто:Where

is the smallest code length that can be assigned to blocks from the group W _{n, k} . In addition, since the words in the group W _{n, k} follow in lexicographic order, the division into W _{n, k, l} and W _{n, k, l + 1} is simple:

,

,

,

,

where n _k denotes the size of the subgroup with shorter codewords. Thus, if the first subgroup has three codewords with a length of 3, and the second subgroup of the same group has one codeword with a length of 4, then n _k (the size of the subgroup with shorter codewords, that is, the first subgroup) is 3. This the example corresponds to subgroups in the group associated with layers 3 and 4 of the coding tree of FIG. 6, in which subgroup 68A has codewords 001, 010, and 011, each of which is three in length, and subgroup 68B has codeword 0001, which is four in length.

The lexicographically smallest codewords in each subgroup can be called basic codewords, for example, as described above with respect to the first aspect of the present invention, and can be represented as:

,

,

,

,

where w _i is the i-th block in the group W _{n, k} . Note that, as indicated above, the remaining codewords in both subgroups can be calculated as follows:

As an illustration, suppose that there is a first subgroup 68A with three codewords of length 3 and a second subgroup 68B with one codeword of length 4, for example, as in the example of layers 3 and 4 of the coding tree of FIG. 6. In this case, if the position of this block is i = 2, then i <n _k (n _k is 3), and the resulting codeword is the applicable base word + i. In this example, the base codeword for the subgroup is 001, and the resulting codeword is 001 + 2 = 011. For levels 3 and 4 of the coding tree in FIG. 6, if the position of the applicable codeword were i≥n _k , then the codeword would be in the second subgroup and would be equal to the base codeword 0000 + in _k , which is 0000 + 4 -3 = 0001.

Basic codewords are defined only by non-empty subgroups, and the number of such subgroups S in the tree constructed for n-bit blocks is within:

n + 1 ≤ S ≤ 2n.

In addition, at the same level there can be many subgroups, and the number of such related subgroups cannot exceed n + 1. For example, at the 10th level of the tree in FIG. 6, there are two subgroups corresponding to codewords 1110 and 1111. However, these subgroups do not belong to the same group. This is a significant difference from algorithms that assign unique basic codewords for each level, but then require a large reordering table of size O (n2 ⁿ ) to work with such codes. In this case, the whole structure has a size of O (n ² ) bits.

In general, this code structure defines groups W and subgroups S. Each of the groups includes code words representing values having the same weight. The code words in each of the groups are ordered in lexicographic order relative to the values represented by the code words. In addition, the first subgroup in each group includes codewords having a first length, and the second subgroup includes codewords having a second length other than the first length. The code structure can be represented by a data structure that the encoder can access to perform variable length encoding. The invention provides for the use of such a code structure in variable-length encoding or decoding, as well as a computer-readable medium containing a data structure that defines such a code structure.

Next, based on the foregoing discussion, a simple algorithm for directly computing block codes will be described. It is assumed that the parameters n _k (0≤k≤n) are available and that level l and the base codeword B _{n, k, l} can be obtained for each non-empty subgroup. Then, the encoding process of the block w can be performed essentially by the following steps:

1. If there is a block w, get its weight k and index I _{n, k} (w).

2. If I _{n, k} (w) <n _k , choose the first subgroup W _{n, k, l} ; otherwise, choose the second subgroup W _{n, k, l + 1} .

3. Then extract the base codeword (B _{n, k, l} or B _{n, k, l + 1} ) for the selected subgroup (W _{n, k, l} or W _{n, k, l + 1} ) and calculate the corresponding code using the following the formula:

(13)

(13)

According to the above formula, if the position i = I _{n, k} (w) of the block w in the selected subgroup (W _{n, k, l} or W _{n, k, l + 1} ) is less than the number n _k blocks in the subgroup, then the code word B _{n , k, l} + i. If the position i of the block w in the selected subgroup (W _{n, k, l} or W _{n, k, l + 1} ) is greater than or equal to the number n _k blocks in the subgroup, then the code word is B _{n, k, l} + in _k .

As described above and shown in the illustration for levels 3 and 4 of the coding tree of FIG. 6, the above process outputs a codeword 011 when the position of a given block code is i = 2 <n _k , and codeword 0001 when a position of a given block code is i = 3≥n _k . In this example, n _k is 3, that is, the number of codewords in the first subgroup 68A for a weight k = 1. The order of position i is determined lexicographically, starting with the base codeword, for example, from 0 to 3 in the case of weight = 1 in the example of FIG. 6. In particular, position 0 corresponds to the base codeword 001, position 1 corresponds to the codeword 010, position 2 corresponds to the codeword 011, all these codewords are in the first subgroup 68A (i <n _k ), and position 3 corresponds to the codeword 0001 in the subgroup 68B (i≥n _k ).

This process can be easily implemented by reordering codewords located at the same level so that they have the same lexicographic order as the input blocks that they represent. For example, the codewords described above are followed in the lexicographical order of input blocks 0001, 0010, 0100 and 1000. Then, the lexicographically smallest codewords in each subgroup, for example, 001 in subgroup 68A or 0001 in subgroup 68B, can be used as basic codewords for the purpose of the above calculation code word. Table 7 below shows the C code that represents an example of the implementation of the process for directly building block codes described above.

Table 7
The process of directly building block codes

In the above C code, the value k indicates the weight of the block w, the value i indicates the position I _{n, k} (w) of the block within the group with weight k, and nk [k] indicates the number of code words in the first subgroup of the group with weight k. If i is greater than or equal to nk [k], then i decreases to adjust the index, and the subgroup is established in the second subgroup (1) for the corresponding weight k. This second subgroup is identified by j = sg [k] [1]. If i is less than nk [k], i does not decrease, and the subgroup is established in the first subgroup (0) for the corresponding weight k. This first subgroup is identified by j = sg [k] [0].

Then a codeword is generated as the sum of the base codeword for the corresponding subgroup j (base [j]) and the value i. In the case of the example of FIG. 6, if the value of i is 2, then the code is the sum of the base codeword 001 for subgroup 68A and the value of i (2), which is 001 + 010 = 011. As follows from the above formula (13), depending on the subgroup, the base codeword is either B _{n, k, l} , or B _{n, k, l + 1} , and the value i is either i or i-nk [k ]. Therefore, the above code generally corresponds to the result given by formula (13). After calculating the code word (code), the length (len) is defined as len [j], which is the code length for the corresponding subgroup, where the code length of the second subgroup is one greater than the code length of the first subgroup. The encoding process then writes the bitstream using the put_bits (code, len, bs) operation, which writes the value of code and len to the bs bitstream. The bitstream is transmitted for decoding by another device. The process returns the weight k to calculate the next codeword.

The above encoding process may include selecting one of the groups based on the weight of the value to be encoded, selecting one of the subgroups based on the lexicographic position of the value to be encoded, relative to the number of code words in the first subgroup of the selected group, selecting one of the code words in the selected subgroup based on the code word for the selected subgroup and the lexicographic position of the value to be encoded, and encoding the value to be encoded using the selected code fishing. The base codewords for each of the subgroups, the position of the codewords within each of the groups, the number of codewords within each of the first subgroups, and the length of the codewords within each of the subgroups can be stored in a data structure that can be accessed by the encoder to support variable length encoding.

In terms of memory, the process shown in Table 7 requires only S base code words (length O (n) bits), n + 1 nk values (length O (n) bits), S code lengths (length O (log n) bits) and 2 (n + 1) indexes of subgroups (length O (log n) bits). An additional memory reduction is possible by storing the increments of the base codewords, which was considered elsewhere in this application. For S = O (n), the entire data structure requires only O (n ² ) bits. In the particular embodiment shown in Table 7, assuming, for example, that n is 20 and S = 32, the size of this data structure becomes 244 bytes. This is much less than 220 words that would be needed to present this code directly in a table. For reasonably short blocks, for example, for n≤12 ... 16, the calculation of weights and indices (functions weight (.) And index (.,.) In the process shown in Table 7) can be a single search. In this case, the entire encoding process may require a maximum of one comparison, two additions and four searches.

For larger blocks, the following well-known combinatorial formula can be used:

,

,

для всех k>n. Для реализации этой формулы можно либо заранее вычислить все биноминальные коэффициенты до уровня n в треугольнике Паскале, либо вычислять их динамически, используя следующие простые тождества:where w _j are the individual bits of the word w, and it is assumed that

for all k> n. To implement this formula, you can either pre-calculate all binomial coefficients up to level n in the Pascal triangle, or dynamically calculate them using the following simple identities:

и

.

and

.

слов (O(n³) битов) памяти и O(n) сложений. Динамическое вычисление коэффициентов требует O(n) сложений, умножений и делений. Однако весь процесс может требовать лишь нескольких регистров и не требовать статистической памяти. Дополнительное обсуждение вопроса сложности вычисления индексов можно найти в работе T. Tjalkens, Implementation cost of the Huffman-Shannon-Fano code, in Proc. Data Compression Conference (DCC'05) (Snowbird, Utah, March 29-31, 2005) 123-132.Implementation based on pre-calculated coefficients requires

words (O (n ³ ) bits) of memory and O (n) additions. Dynamic coefficient calculation requires O (n) additions, multiplications and divisions. However, the entire process may require only a few registers and not require statistical memory. An additional discussion of the difficulty of computing indexes can be found in T. Tjalkens, Implementation cost of the Huffman-Shannon-Fano code, in Proc. Data Compression Conference (DCC'05) (Snowbird, Utah, March 29-31, 2005) 123-132.

The following describes an example of a design of a decoder that implements the above methods. Like the encoding process described above, the decoding process can use the parameters n _k , the base codewords and the associated lengths. For convenience, the following discussion uses left-aligned variants of the values of the basic codewords:

,

,

where T is the length of the machine word (T> max l). In this case, an exemplary decoding process can be described as follows:

меньше, чем T бит в потоке битов.1. Find the uppermost subgroup for which

less than T bits in the bitstream.

2. Decode the block index I _{n, k} (w) based on the above formula (13).

3. Get the restored block using its weight k and index.

Table 8 below shows the C code representing an example implementation of the above decoding process.

Table 8
Decoding Code Blocks

The exemplary decoding process shown in Table 8 uses the left-aligned base codewords lj_base [S]. In operation, the decoding process receives the contents of val from the bitstream buffer and identifies a subgroup in the encoding tree having a base codeword corresponding to the contents of val of the bitstream buffer. For example, a process continues to cross, going down, subgroups at various levels in the coding tree until the base codewords are larger than the received codeword. However, when the process reaches a subgroup with a base codeword that is less than or equal to val, that subgroup is selected. After finding the appropriate subgroup, the process determines the code length for the subgroup and then scrolls the bitstream by this length to skip the decoded bits and isolate the codeword. The decoding process determines the index position i of the codeword within a subgroup by subtracting the value of the base codeword from the contents of the bitstream buffer.

If, for example, the codeword is 011, and the base codeword is 010, then the result of this difference is 2, which means that the codeword is at position 2 from possible positions 0, 1, and 2 in the subgroup. For the case of a register 32 bits wide, this difference can be shifted to the right by 32 minus the length len of the code. The decoding process then extracts the corresponding weight k and subgroup index j and restores the index i. The process then generates the ith word in the selected group as a codeword and returns the weight k. The expression kj [j] .j returns the weight of the subgroup, and the expression kj [j] .j returns the index of the subgroup as either 0 or 1, which indicates either the first subgroup (0) or the second subgroup (1) for a given weight . If the second subgroup is chosen in such a way that j = 1, then the index i is adjusted by adding the value nk [k]. Or index i is not adjusted if the first subgroup is selected. The word () function returns the ith word to the group n, k as a decoded word value, for example, using the above formula (13).

In general, the encoder can perform variable length coding according to the above code structure, the code structure defining groups and subgroups. Once again, each of the groups includes codewords representing values with the same weights. The code words in each of the groups are ordered in lexicographic order relative to the values represented by the code words. In addition, the first subgroup in each group includes codewords having a first length, and the second subgroup includes codewords having a second length other than the first length.

The code structure can be represented by a data structure that can be accessed by an encoder or decoder to perform variable length encoding. As described above, the data structure can determine the basic codewords for each of the subgroups, the position of the codewords within each of the groups, the number of codewords within each of the first subgroups, and the length of the codewords within each of the subgroups. This data structure can be stored in memory associated with one encoder from a video encoder, an image encoder, an audio encoder, a voice encoder, a video data decoder, an image decoder and an audio decoder or voice decoder, and can be accessed to support encoding operations as necessary.

As described above, a decoder, such as statistical decoding unit 52, may select during a top-down search in the encoding tree by selecting the first of the subgroups with a base codeword that is less than or equal to the codeword to be decoded. The decoder can then determine the position of the codeword to be decoded within the selected subgroup, that is, the index of the subgroup, based on the difference between the codeword to be decoded and the base codeword for the selected subgroup. The decoder determines the weight of the value represented by the codeword to be decoded based on the group in which the selected subgroup is located, and determines the position, that is, the index of the group, code word within the group in which the selected subgroup is based, based on whether the selected subgroup is the first subgroup or the second subgroup for the group. The decoder then selects one of the values based on the weight of the value represented by the codeword to be decoded and the position of the codeword within the group in which the selected subgroup is located, and decodes the codeword to be decoded using the selected value. This value may correspond, for example, to one of the block codes in table 6.

The code structure and data structure provided in accordance with the present aspect of the invention can maintain efficiency in terms of excessive computational overhead and processing time. For example, the example decoding process in Table 8 requires a subgroup from 1 to S to compare and search, one or two additions, one shift operation, one additional comparison operation, and three additional search operations to find a subgroup. The number of steps required to find a subgroup can be further reduced by placing the base code words in the binary search tree or by using the additional search table, but in both cases due to the additional use of memory.

At the end of the decoding process described above, the weight k and index I _{n, k} (w) for the codeword are converted to actual values (for example, using the word () function in table 8). If the blocks are reasonably short, this can be achieved simply by a search operation. Otherwise, the word can be synthesized by using the numbering formula, for example, as described in DA Huffman, A method for the construction of minimum-redundancy codes. Proc. IRE, 40 (Sept. 1952) 1098-1101. In terms of complexity, this process is similar to calculating an index in an encoder.

Using the above encoding and decoding processes, it is possible to define a system for adaptively encoding and decoding data blocks. In this example, it is assumed that input blocks can be encoded under the following conditions:

1. There is no context, that is, universal code is implemented.

2. The context is given by one previously observed block, that is, t = n.

3. The context is given by two previously observed blocks, that is, t = 2n.

Instead of using actual blocks as a context, it is enough (due to lack of memory at the source) to use their weights. This means that an array of t + 1 code structures indexed by their weight s is provided for t-bit samples. To further save space, one can use the symmetry of KT distributions with respect to s and k. In particular, the process can replace s = t-s and flip bits (that is, force k = n-k) every time s> t / 2. Thus, it is only necessary to define t / 2 + 1 tables. In this example, the total amount of memory required by the adaptive code becomes 1 + 0 + 1 + n + 1 = 1.5n + 3 tables. Special cases of estimating the amount of memory for block sizes n = 8 ... 20 are shown below in table 9.

Table 9
Estimates of memory usage (in bytes) for various block sizes n Max. t Max. S Size of one table Tables for all contexts 8
12
16
twenty 16
24
32
40 fourteen
19
25
29th 102
140
184
216 1530
2940
4986
7128

The above tables were generated using densities based on CT scores and using a modified Huffman code algorithm in accordance with the present invention. The following is an example of computer code for a program that implements the adaptive block encoder described in the present invention.

The following describes the experimental results of evaluating the characteristics of the adaptive coding process described in this application with a block size of n = 16 and are compared with other known algorithms. In particular, the adaptive coding process is compared with the Q-encoder algorithm described by WB Pennebaker, JL Mitchell, GG Langdon, Jr., RB Arps, An overview of the basic principles of the Q-Coder adaptive binary arithmetic coder, IBM J. Res. Dev., 32 (6) (1988) 717, which is used in the JBIG image encoding standard, and with the CABAC algorithm described by D. Marpe, H. Schwartz, and T. Wiegand, Context-Based Adaptive Binary Arithmetic Coding in the H.264 / AVC video compression standard, IEEE Trans, on CSVT, 13 (7): 620-636, July 2003, which is used in ITU-T H.264 / MPEG AVC standards.

For testing, computer-generated bit sequences were used to simulate output sequences from a Bernoulli binary source with probability p. The length of such sequences is in the range from 16 to 1024, and for each length Q = 1,000,000 samples from such sequences were generated. The relative degree of redundancy was calculated by the formula:

For the adaptive coding process described in this application, the following context structure was used:

1) the first 16-bit block is encoded without context (universal code);

2) the second block is encoded using one of its contexts (code with t = 16); and

3) the third and all subsequent blocks are encoded using the two previous blocks in the sequence as contexts (code based on samples with t = 32).

7A and 7B are graphs illustrating a comparison of the degrees of redundancy of an adaptive block code with the Q-encoder and CABAC methods at various p values. FIG. 7A shows the results for p = 0.1, and FIG. 7B shows the results for p = 0.5. On Fig is a graph illustrating the sensitivity of redundancy to the asymmetry of the source data for adaptive block code, methods of Q-encoder b CABAC. The results of the experimental study are shown in Figs. 7A, 7B and 8. In Figs. 7A and 7B, the number of encoded bits is plotted on the X-axis, and the ratio (average code length - entropy) / entropy is plotted on the Y-axis.

In Fig. 8, the x-axis represents the probability, and the y-axis represents the ratio (average code length - entropy) / entropy. Each graph in FIGS. 7A, 7B, and 8 contains graphs for a Q-encoder, CABAC, and adaptive coding with corresponding notation. From the experimental results, it can be noted that the adaptive code described in this application may have a significantly higher convergence rate compared to the Q-encoder and CABAC algorithms. The adaptive coding process surpasses the Q encoder and CABAC algorithms for shorter sequences, and becomes comparable to the Q encoder and CABAC algorithms when the total length of the encoded bits approaches 1024. In addition, from Fig. 8, it follows that after 160 encoded bits ( or, by that time, 16-bit blocks) the adaptive coding process can provide lower redundancy compared to the Q-encoder and CABAC algorithms. This pattern is consistent with Theorem 1 above.

FIG. 9 is a flowchart illustrating a method for constructing variable length coding with efficient use of memory for monotone distributions in accordance with a first general aspect of the present invention. The method can be implemented by a processor associated with an encoder, decoder, or other device for constructing codes for use by the statistical encoding module 46 and the statistical decoding module 52, as shown in FIGS. 2 and 3, and can support compression and encoding of a wide variety of data, including, in particular, video data, image data, voice and sound data. Such a processor can be provided, for example, as part of an encoder or decoder or in a general-purpose computing system, for example, to create a data structure that defines the attributes of the code structure used in variable-length encoding.

As shown in FIG. 9, the processor obtains an alphabet of input characters to be encoded (70). Symbols have corresponding weights indicating the likelihood or frequency of the presence or use of symbols in a given data set or sequence. After determining the corresponding symbol weights (72), the process assigns indices to the symbols based on their weights (74) and assigns codes to the symbols based on the indices and the lexicographic order of the symbols (76). Therefore, characters having the same weights can be ordered in lexicographic order to facilitate the encoding methods described in this application.

Codes can be organized according to the code structure represented by the binary encoding tree. The processor identifies the base codeword for each level in the encoding tree (78). The base codeword is the smallest codeword at a given level in the tree and lexicographically corresponds to the earliest character among the characters at that level in the tree. To ensure a compact data structure, the processor removes a fixed number B of the initial bits from the base codewords to obtain partial base codewords (80). The base codewords can be defined as left-aligned codewords, and the leading bits can be M leading bits when passing from right to left in left-aligned codewords. In some embodiments, the number of starting bits to be deleted may be 8. In other embodiments, the number of starting bits to be deleted may be less than or greater than 8.

For many levels of the coding tree, the M leading bits are zeros. However, at some levels, the start bits may form all or part of the base code for the corresponding level in the tree. At these selected levels, the processor generates skip indicators (82). The skip indices provide the decoder with a command to scroll the bitstream by B bits so as not to lose the base code after deleting the initial B bits. A process can store in the data structure the resulting partial base codewords, lengths associated with codewords at the appropriate levels of the coding tree, and one or more skip indicators that indicate when the bitstream should be scrolled to B bits to prevent loss of the base codeword , which is at least partially in the initial B bits (84). This data structure can be provided to the statistical encoding module 46 and the statistical decoding module 52. The data structure can take various forms, including one-dimensional, more than one-dimensional and multidimensional search tables, link lists, binary trees, base trees, structureless files, or the like.

FIG. 10 is a flowchart illustrating a method of encoding symbols using variable-length codes constructed according to the method of FIG. 9 in accordance with a first general aspect of the present invention. As shown in FIG. 10, the statistical coding unit 46 receives a symbol (86), determines an index for a symbol (87), and compares the symbol index with an offset table to identify the corresponding level in the encoding tree. In particular, statistical coding unit 46 determines whether a character index is greater than or equal to an offset for a given tree level (88). A character index is a character order among other characters, ranked in order of weight, with characters of equal weight being ordered in lexicographic order corresponding to the character alphabet. The offset is the difference between the length of the code or codes for the corresponding tree level and the maximum code length. If in the tree of FIG. 4 the maximum code length is, for example, 16 and the code length at tree level 3 is 3, then the corresponding offset of the base codeword is 12.

If the symbol index does not exceed the offset for the current level of the tree, the statistical coding unit 46 moves down to the next level (90) of the coding tree in the downward search process and repeats the comparison of the symbol index with the offset for this next level (88). When the statistical encoding unit 46 determines that the index of the symbol is greater than or equal to the offset for a particular level of the encoding tree (88), the statistical encoding unit calculates the corresponding code word for that symbol. In particular, statistical encoding module 46 sets the codeword for a symbol to the sum of the base codeword for the current tree level and the difference between the symbol index and the offset for that level (92).

If we take the tree in FIG. 4 as an example, then if the symbol index is 14, the statistical coding module 46 determines that the code for the symbol is at tree level 3, because 14 is greater than the offset 12 assigned to the base codeword for this level. Then, the statistical coding unit 46 calculates the codeword as a base codeword (001) + (character index (14) - offset (12)), i.e. 001 + 2 = 001 + 010 = 011. After setting the code for the received character (92 ), the statistical encoding module 46 provides a codeword to the bitstream (94) for transmission, for example, to a receiver with the statistical decoding module 52. Then, statistical encoding unit 46 repeats the process for the next character in the corresponding data sequence.

11 is a flowchart illustrating a method for decoding variable length codes constructed according to the method of FIG. 9 in accordance with a first general aspect of the present invention. The method depicted in FIG. 11 may be implemented using an algorithm identical or similar to the algorithm shown in Table 5. In the embodiment, the codes may be received by the statistical decoding module 52, the codes being encoded by the statistical encoding module 46, as described with reference to figure 10. The method of FIG. 11 can use the memory efficient decoding methods described herein and can take advantage of the compact data structure by which such codes can be constructed. As shown in FIG. 11, the statistical decoding unit 52 receives a codeword from an incoming bitstream (96). The codeword can be obtained from a fixed width of W bits extracted from the bitstream buffer. The codeword can be left justified or converted to left justified formats, and the width W can be reduced by removing B leading bits from the codeword from right to left.

The statistical decoding unit 52 compares the codeword with the base codeword for various levels of the coding tree, starting from the top level and moving to deeper levels in the downward search process until the corresponding base codeword is found. In particular, the statistical decoding unit 52 may determine whether the base codeword for the current tree level is less than or equal to the codeword (98). If not, the statistical decoding unit continues to move down to the next level of the tree (100) and repeats the comparison (98) for the base codeword associated with the next level. However, after moving to the next level (100), the statistical decoding module 52 determines whether the skip indicator (102) is associated with the current level. If connected, then before moving on to the next level (100), the statistical decoding module 52 scrolls the bitstream buffer by a fixed step (104). In particular, the statistical decoding unit 52 may scroll the bitstream buffer by M bits so that the codeword is not lost when the initial M bits are skipped. If there is no skip sign (102), the statistical decoding unit 52 simply proceeds to the next level (100).

In both cases, the statistical decoding unit 52 again compares the codeword with the base codeword for the current level (98). When the statistical decoding unit 52 finds a level at which the base codeword is less than or equal to the codeword (98), the statistical decoding unit 52 determines the residual length of the base codewords at the corresponding level (106) and scrolls the bitstream by the residual length (108). Then, the statistical decoding unit 52 calculates a symbol associated with the codeword (110) based on the offset for the level and the difference between the base codeword and the decoded codeword.

If we take the tree in FIG. 4 as an example, then if the codeword is 0110000000000000, then the partial, truncated codeword with the missing 8 leading bits is 01100000. In this case, the statistical decoding module 52 determines that the partial base codeword is at level 3 (00100000 ) is less than or equal to the codeword, and identifies a residual length of 3. The statistical decoding unit 52 scrolls the bitstream 3 bits forward to obtain the next codeword. In addition, the statistical decoding unit 52 computes the symbol for the codeword by adding an offset for level 3 to the difference between the codeword in the data stream buffer and the base codeword for that level. For example, statistical decoding unit 52 computes the code as an offset [3] = 12 plus codeword 01100000 minus 00100000, which is equivalent to 12 plus 2 = 14. In this case, 14 is the index of the character represented by code 011.

The method depicted in FIG. 11 can take advantage of a very compact data structure and very efficient use of memory, as described elsewhere in this application. As a result, when implementing this method, the statistical decoding module 52 can provide increased efficiency, including reduced processing costs, reduced memory usage and increased processing speed, all of which it is desirable to have in video decoding devices and in other devices capable of unpacking and decoding data.

12 is a flowchart illustrating a method for constructing adaptive block codes in accordance with a second general aspect of the present invention. The method of FIG. 12 can be implemented in a processor located in an encoder, or in a general-purpose processor to support efficient, direct construction of adaptive block codes. As shown in FIG. 12, the processor obtains a set of words (112) to be encoded. The data structure, which is the resulting code structure, can be stored in memory inside the encoder, decoder, or both. Words can be blocks of binary codes. After determining the word weights (114), the processor assigns groups of code words to words based on the weights (116). The codeword groups include codewords for words with the same weight k and can span two adjacent levels of the coding tree, for example, as shown in FIG. 6.

As shown further in FIG. 12, the processor assigns subgroups to the words of one group based on word lengths (118). In particular, a group may include a first subgroup and a second subgroup. The first subgroup contains one or more code words having the same length and the same weight. Similarly, the second subgroup contains one or more code words having the same length and the same weight. However, the codeword length in the first subgroup is less than the codeword length in the second subgroup. Thus, each subgroup includes codewords that have the same weight and are at the same level in the encoding tree.

The processor identifies the base codeword for each subgroup (120). A base codeword is the smallest codeword in a subgroup. In the example of FIG. 6, the base codeword for subgroup 68A is 001. However, subgroup 68 further includes codewords 010 and 011 in addition to base codeword 001. In this example, codewords in the subgroup are ordered in the lexicographic word order that they represent, so that codes can be easily and directly calculated with a relatively small amount of computation.

The number of elements in the first subgroup of each group can be used to calculate codewords. For this purpose, the processor stores the number of elements in the first subgroup of each group (122), and also stores the group index display (124), the subgroup code length display (126) and the subgroup base code word display (128). The group index display may identify the position of the codeword for the word in the group in which the codeword is located. The display of the length of the codes of a subgroup can identify the length of the codes within a specific subgroup. The display of the base codeword of a subgroup can identify the base codeword associated with each subgroup. In general, a code can be constructed from a basic codeword for a particular subgroup if the index of the word in the group is known. The stored information may be stored in a data structure that can be accessed by an encoder, decoder, or both.

FIG. 13 is a flowchart illustrating a method of encoding blocks using variable-length codes constructed according to the method of FIG. 12 and in accordance with a second general aspect of the present invention. The method of FIG. 13 may be implemented, for example, in a statistical coding unit, such as statistical encoding unit 46 in FIG. 2. The method shown in Fig. 13 can be implemented using an algorithm identical or similar to the algorithm shown in Table 7. As shown in Fig. 13, to encode this word, the statistical encoding module 46 receives its weight (130) and the group index ( 132). Using the word weight, the statistical coding unit 46 determines the group index for the word (132) within the corresponding coding tree. The index I _{n, k} (w) of the group defines the index (position) of the word w in the group W _{n, k} , which stores the words in lexicographic order.

The statistical coding unit 46 compares the group index with the number of elements n _k in the first subgroup of the group in which the code word for the input word is located. In particular, statistical coding unit 46 determines whether the group index is greater than or equal to the number of elements in the first subgroup (134). If it is, statistical coding unit 46 selects a second subgroup (138) in the group, that is, subgroup 1, and decreases the value of the group index (140). In particular, the value of the group index decreases by the number of elements in the first subgroup n _k . If the group index is not greater than or equal to the number of elements in the first subgroup (134), the statistical coding unit 46 selects the first subgroup (136), that is, the subgroup 0. Each subgroup has its own base codeword. The statistical coding unit 46 obtains the base codeword for the selected subgroup (142) and calculates the codeword based on the sum of the base codeword and the group index value (144).

If we refer to the example of the encoding tree in Fig. 6 as an illustration, then assuming that, for example, the weight of the word to be encoded is 2, the index value of the group is 2, and the number of elements in the first subgroup is 3, which corresponds to the group at the levels 3 and 4 of the encoding tree, then in this case the statistical encoding module 46 selects the first subgroup (subgroup 0), since the index value of the group (2) is less than the number of elements (3) in the first subgroup. As a result, the base codeword is 001. For encoding the word, the statistical coding unit 46 adds the group 2 index value to the base codeword 001, which gives the codeword 011.

If for the same group the group index value was 3, then the statistical coding module 46 would select the second subgroup (subgroup 1). However, statistical coding unit 46 would reduce the group index value by the number n _k elements in the first subgroup (subgroup 0). In this case, the group index value of 3 would be reduced by 3, which is zero, and the codeword would be calculated as the base codeword 0001 for the second subgroup plus the group index value of 0, which gives the codeword 0001.

In addition to calculating the codeword for the input word, the statistical coding module 46 may obtain the length of the codes in the selected subgroup (146). In the above example, for subgroup 0 at level 3, the length of the codes would be 3. The statistical encoding unit outputs the calculated codeword and the length of the subgroup code to the bitstream for storage and / or transmission to another device, such as a decoding device including a statistical decoding unit such as statistical decoding unit 52.

FIG. 14 is a flowchart illustrating a method for decoding variable length codes constructed according to the methods of FIG. 12 and FIG. 13 and in accordance with a second general aspect of the present invention. The method depicted in FIG. 14 may be implemented using an algorithm identical or similar to the algorithm shown in Table 8. Variable-length codes may be obtained from an encoder, such as an encoder, including a statistical encoding unit 46. Variable length codes can be obtained and decoded by the statistical decoding unit 52. As shown in FIG. 14, the statistical decoding unit 52 obtains a codeword from an incoming bitstream (150) and compares the codeword with the base codeword of the subgroup. In particular, statistical decoding unit 52 may search for a corresponding coding tree for a left-aligned subgroup base word that is less than or equal to the contents of the code word obtained from the bitstream buffer (152).

If the base codeword of the subgroup in the subgroup at a given level of the tree is not less than or equal to the codeword (152), then the statistical decoding module 52 proceeds to the next subgroup at the next level (154) of the tree and repeats the comparison. This process continues on a phased basis until the base codeword is greater than or equal to the codeword received from the bitstream, that is, until the statistical decoding unit 52 reaches a level at which the base codeword of the subgroup is less than or equals code word. When comparing the codeword and the base codewords, the statistical decoding unit 52 may use partial, incremental values of the base codewords to further reduce memory in accordance with the first aspect of the present invention. In particular, a few leading bits can be skipped to improve memory efficiency, as described above.

When the level of the encoding tree is reached at which the base codeword of the subgroup is less than or equal to the codeword, the statistical decoding unit 52 determines the code length for the subgroup (156) and scrolls the bitstream by this length to skip the decoded bits and isolate the codeword. The statistical decoding unit 52 may determine the position of the codeword in a subgroup using the base codeword (158). For example, statistical decoding unit 52 may subtract the codeword of the bitstream from the base codeword to obtain a position difference between the codeword and the base codeword.

If we take the coding tree of FIG. 6 as an example, then if the incoming codeword is 0110000000000000, the statistical decoding unit 52 identifies the base codeword 0010000000000000 as the highest base word, the codeword of the subgroup, which is less than or equal to the codeword. This base codeword is associated with the first subgroup in the group at levels 3 and 4. After determining the code length (3) of the subgroup associated with the base codeword, the statistical decoding unit 52 can scroll the bitstream to skip the decoded bits. The statistical decoding unit 52 may determine the group index for the codeword by subtracting the base codeword from the codeword obtained from the bitstream. In this example, 011 minus 001 gives 010, which gives a group index value of 2.

The statistical decoding unit 52 may also determine the weight of the codeword (160), for example, based on the level of the subgroup in the encoding tree. In addition, the statistical decoding unit 52 may determine the index of the subgroup (162), that is, the index of the selected subgroup in the encoding tree. Using the weight, position and index of the subgroup, the statistical decoding unit 52 determines the index of the word (164), thereby decoding the codeword obtained from the data stream to obtain the word represented by the codeword. Once again, in some embodiments, the decoding method performed by the statistical decoding unit 52 may correspond to the process shown in Table 8.

Specialists in the art should understand that information and signals can be represented using a variety of technologies and methods. For example, data, instructions, commands, information, signals, bits, symbols and elements that may be mentioned in the above description may be represented by voltage, current, electromagnetic waves, magnetic fields or particles, optical fields or particles, or any combination of the above.

Those skilled in the art will also appreciate that the various logical blocks, modules, circuits, and algorithm steps described by way of illustration described in connection with the disclosed embodiments may be implemented as electronic hardware, computer software, or a combination of both funds. To more clearly illustrate this interchangeability of the hardware and software provided by way of illustration, various components, blocks, modules, circuits, and steps have been described above in terms of their functionality. Whether such functionality will be implemented in the form of hardware or software depends on the specific application and design restrictions imposed by the system as a whole. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be construed as causing a departure from the scope of the present invention.

The described methods can be implemented in hardware, software, hardware-software, or any combination of the above. Such methods can be implemented in any of a variety of devices, such as general-purpose computers, wireless telephone devices, or integrated circuit devices having many fields of application, including applications in wireless telephone devices and other devices. Any characteristics described in the form of modules or components can be implemented together in an integrated logical device or separately as separate, but interacting logical devices. In the case of a software implementation, these methods can be implemented, at least in part, by a computer-readable medium for storing data containing program code including instructions that, when executed, execute one or more of the above methods. Computer-readable storage media may be part of a computer program product. A computer-readable medium may comprise random access memory (RAM), such as synchronous dynamic random access memory (SDRAM), read-only memory (ROM), non-volatile random access memory (NVRAM), electrically erasable programmable read-only memory (EEPROM), flash memory, magnetic or optical media for storing data, etc. In addition to or instead, these methods can be implemented, at least in part, by a computer-readable transmission medium that transfers or transmits program code in the form of instructions or data structures which may be available, readable and / or executable computer, such as propagate signals or waves.

The program code may be executed by one or more processors, such as one or more digital signal processors (DPS), general purpose microprocessors, specialized integrated circuits (ASICs), user programmable large integrated circuits (FPGAs), or other equivalent integrated or discrete logic circuits. A general purpose processor may be a microprocessor, but in the alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. A processor can also be implemented as a combination of computing devices, for example, in the form of a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in combination with a DSP core, and any other such configuration. Accordingly, the term "processor" in this application may refer to any of the above structures or to any other structure suitable for implementing the methods described in the application. In addition, in some aspects, the described functionality may be provided within specialized software modules or hardware modules capable of encoding or decoding, or may be integrated into video codec (CODEC).

Various embodiments of the invention have been described. These and other embodiments of the invention are within the scope of the following claims.

Claims (83)

1. A method of creating data for encoding code words of variable length, comprising the steps of:
generating partial values of the base codewords, the base codewords corresponding to the levels of the coding tree defining canonical codewords of variable length, and the partial values being based on the shift of the base codewords by a fixed number of bits;
generating a skip indicator instructing the decoder to skip a fixed number of bits in the bit stream to be decoded before moving to a selected level of the encoding tree; and
store partial values and the skip rate in the data structure in memory. 2. The method according to claim 1, additionally containing stages in which:
generating values represented by basic codewords;
generate lengths of partial values of the base codewords and store the values and lengths in the data structure in memory. 3. The method according to claim 1, further comprising storing the data structure in the memory of one of the video decoder, image decoder and sound decoder or voice decoder. 4. The method according to claim 1, in which some of the levels of the coding tree include codewords arranged in lexicographic order relative to the values represented by the codewords, and each of the base codewords is the lexicographically smallest codeword at the corresponding level in the coding tree. 5. The method according to claim 1, in which the partial values of the base code words are the removal of a fixed number of leading bits from the base code words. 6. The method according to claim 5, in which, in the absence of said offset, the base codeword at a selected tree level extends, at least in part, to a remote number of start bits. 7. The method according to claim 1, further comprising the step of decoding one of the codewords from the bitstream with the codewords using a stored data structure, the decoding comprising the steps of:
searching for coding tree levels for a selected value from partial values of the base codewords that is less than or equal to the codeword from the bitstream with the codewords;
skip a fixed number of bits in the bitstream with the code words before moving to the selected level of the coding tree in response to the skip rate;
one of the plurality of values corresponding to the codeword is calculated based on the difference between the selected value from the partial values of the base codewords that is less than or equal to the codeword and the codeword and the index of the selected value from the partial values of the base codewords that is less than or equal to the codeword . 8. The method according to claim 7, further comprising the step of presenting the output to the user based at least in part on the calculated value. 9. The method of claim 1, further comprising encoding the values with codewords consistent with the coding tree, the values represented by codewords representing at least one of video data, image data, voice data, or audio data. 10. A device for creating data for encoding codewords of variable length, containing:
means for generating partial values of the base codewords, wherein the base codewords correspond to levels of a coding tree defining canonical codewords of variable length, and wherein the partial values are based on shifting the base codewords by a fixed number of bits;
means for generating a skip indicator instructing the decoder to skip a fixed number of bits in the bit stream to be decoded before moving to a selected level of the encoding tree; and
means for storing partial values and a skip rate in a data structure in memory. 11. The device according to claim 10, further comprising:
means for generating values represented by basic codewords;
means for generating lengths of partial values of the base codewords, and
means for storing values and lengths in a data structure in memory. 12. The device of claim 10, further comprising means for storing the data structure in the memory of one of the video decoder, image decoder, audio decoder, or voice decoder. 13. The device according to claim 10, in which some of the levels of the coding tree include codewords arranged in lexicographic order relative to the values represented by the codewords, and each of the basic codewords is the lexicographically smallest codeword at the corresponding level in the coding tree. 14. The device according to claim 10, in which the partial values of the base code words are the removal of a fixed number of initial bits from the base code words. 15. The device according to 14, in which, in the absence of said offset, the base codeword at a selected tree level extends, at least in part, to a remote number of start bits. 16. The device of claim 10, further comprising means for decoding one of the codewords from the bitstream with the codewords using the stored data structure, the decoding means comprising:
means for searching the levels of the coding tree for a selected value from partial values of the base codewords that is less than or equal to the codeword from the bitstream with codewords;
means for skipping a fixed number of bits in the bitstream with codewords before moving to a selected level of the coding tree in response to a skip rate;
means for calculating one of a plurality of values corresponding to the codeword based on the difference between the selected value from the partial values of the base codewords that is less than or equal to the codeword and the codeword and the index of the selected value from the partial values of the base codewords that is less than or equal to the codeword a word. 17. The device according to clause 16, further comprising means for presenting the output to the user based at least in part on the calculated value. 18. The device of claim 10, further comprising means for encoding the values with code words consistent with the coding tree, the values represented by code words representing at least one of video data, image data, voice data, or audio data. 19. A material machine-readable medium containing a data structure storing partial values of the base codewords, the base codewords corresponding to the levels of the coding tree defining canonical codewords of variable length, and the partial values being based on the shift of the base codewords by a fixed number of bits, and an indicator a pass indicating the decoder to skip a fixed number of bits in the bitstream to be decoded before moving to a selected level of the encoding tree. 20. The computer-readable medium of claim 19, wherein the data structure further stores values represented by basic codewords and lengths of partial values of the basic codewords. 21. The computer-readable medium of claim 19, wherein some of the levels of the coding tree include codewords arranged in lexicographic order relative to the values represented by codewords, and each of the base codewords is the lexicographically smallest codeword at a corresponding level in the coding tree . 22. The machine-readable medium of claim 19, wherein the partial values of the base codewords are the removal of a fixed number of start bits from the base codewords. 23. The computer-readable medium of claim 22, wherein in the absence of said offset, the base codeword at a selected tree level extends, at least in part, to a remote number of start bits. 24. The computer-readable medium of claim 22, wherein the values represented by code words represent at least one of video data, image data, voice data, or audio data. 25. A device for creating data for encoding codewords of variable length, comprising:
a processor configured to generate partial values of the base codewords, the base codewords corresponding to the levels of the coding tree defining variable-length codewords, and the partial values being based on the shift of the base codewords by a fixed number of bits, and the generation of a skip indicator indicating the decoder to skip a fixed number of bits in the bitstream to be decoded before moving to a selected level of the encoding tree; and
a memory that stores partial values and a skip rate in the data structure. 26. The device according A.25, in which the processor generates the values represented by the basic code words, and generates the lengths of partial values of the basic code words, and the memory stores the values and lengths in the data structure. 27. The device according A.25, in which the device contains one of the video decoder, image decoder and sound decoder or voice decoder. 28. The device according A.25, in which some of the levels of the encoding tree include codewords located in lexicographic order relative to the values represented by code words, and each of the base code words is the lexicographically smallest code word at the corresponding level in the encoding tree. 29. The device according A.25, in which the partial values of the base code words are the removal of a fixed number of initial bits from the base code words. 30. The device according to clause 29, in which, in the absence of said offset, the base codeword at a selected tree level extends, at least in part, to a remote number of start bits. 31. The device according A.25, further containing a decoder that decodes one of the code words from the bitstream with code words using the stored data structure, the decoder:
looking for coding tree levels for the selected value from the partial values of the base codewords that is less than or equal to the codeword from the bitstream with the codewords;
skips a fixed number of bits in the bitstream with code words before moving to the selected level of the coding tree in response to the skip rate;
calculates one value from the set of values corresponding to the codeword based on the difference between the selected value from the partial values of the base codewords that is less than or equal to the codeword and the codeword and the index of the selected value from the partial values of the base codewords that is less than or equal to the codeword a word. 32. The device according to p, optionally containing an output device that presents output to the user based at least in part on the calculated value. 33. The device according A.25, in which the values represented by code words represent at least one of the video data, image data, voice data or audio data. 34. The device according A.25, optionally containing a wireless receiver that receives code words from the encoder wirelessly. 35. A decoding device comprising:
a memory storing a data structure containing partial values of the base codewords, wherein the base codewords correspond to the levels of a coding tree defining canonical codewords of variable length, the partial values being based on the shift of the base codewords by a fixed number of bits, and a skip rate indicating the decoder to skip a fixed number of bits in the bitstream to be decoded before moving to a selected level of the encoding tree; and
a decoder that accesses the memory to decode one of the codewords from the bitstream based on the partial values and the skip rate in the stored data structure. 36. The device according to clause 35, in which the decoder:
searches for coding tree levels for a selected value from partial values of the base codewords, which is less than or equal to the codeword from the bitstream with codewords,
skips a fixed number of bits in the bitstream with code words before moving to the selected level of the coding tree in response to the skip rate; and
calculates one value from the set of values corresponding to the codeword based on the difference between the selected value from the partial values of the base codewords that is less than or equal to the codeword and the codeword and the index of the selected value from the partial values of the base codewords that is less than or equal to the codeword a word. 37. The device according to clause 36, further comprising an output device that presents output to the user based at least in part on the calculated value. 38. The device according to clause 37, in which the data structure contains the values represented by the base code words, and the lengths of the partial values of the base code words. 39. The device according to clause 35, in which the decoder contains a multimedia decoding device, and the decoder includes one of a video decoder, image decoder and sound decoder or voice decoder. 40. The device according to clause 35, in which some of the levels of the encoding tree include codewords located in lexicographic order relative to the values represented by code words, and each of the basic code words is the lexicographically smallest code word at the corresponding level in the encoding tree. 41. The device according to clause 35, in which the partial values of the base code words are the removal of a fixed number of initial bits from the base code words. 42. The device according to paragraph 41, in which, in the absence of said offset, the base codeword at a selected tree level extends, at least in part, to a remote number of start bits. 43. The device according to clause 35, further comprising a receiver for receiving codewords from the encoder wirelessly. 44. A decoding method, comprising the steps of:
referring to the data structure stored in the memory, the data structure containing partial values of the base codewords, the base codewords corresponding to the levels of the coding tree defining codewords of variable length, the partial values being based on the shift of the base codewords by a fixed number of bits, and the exponent a pass instructing the decoder to skip a fixed number of bits in the bit stream to be decoded before moving to a selected level of the encoding tree; and
decode one of the code words from the bitstream based on the partial values and the skip rate in the stored data structure. 45. The method according to item 44, in which the decoding comprises the steps of:
searching for coding tree levels for a selected value from partial values of the base codewords that is less than or equal to the codeword from the bitstream with the codewords;
skip a fixed number of bits in the bitstream with the code words before moving to the selected level of the coding tree in response to the skip rate; and
one of the plurality of values corresponding to the codeword is calculated based on the difference between the selected value from the partial values of the base codewords that is less than or equal to the codeword and the codeword and the index of the selected value from the partial values of the base codewords that is less than or equal to the codeword . 46. The method according to item 44, further comprising the step of presenting the output to the user based at least in part on the calculated value. 47. The method of claim 44, wherein the data structure comprises values represented by basic codewords and lengths of partial values of the basic codewords. 48. The method according to item 44, in which the decoding comprises the step of decoding the code words to obtain values representing at least one of the video data, image data, audio data and voice data. 49. The method according to item 44, in which some of the levels of the coding tree include codewords arranged in lexicographic order relative to the values represented by codewords, and each of the basic codewords is the lexicographically smallest codeword at the corresponding level in the coding tree. 50. The method according to item 44, in which the partial values of the base code words are the removal of a fixed number of initial bits from the base code words. 51. The method of claim 50, wherein, in the absence of said offset, the base codeword at a selected tree level extends, at least in part, to a remote number of start bits. 52. The method according to item 44, further comprising the step of receiving the code words from the encoder wirelessly. 53. A decoding device comprising:
means for storing a data structure containing partial values of the base codewords, wherein the base codewords correspond to the levels of a coding tree defining variable-length codewords, the partial values being based on the shift of the base codewords by a fixed number of bits, and a skip rate instructing the decoder to skip the fixed number bits in the bitstream to be decoded before moving to the selected level of the encoding tree; and
means for decoding one of the codewords from the bitstream based on partial values and a skip rate in the stored data structure. 54. The device according to item 53, in which the decoding tool comprises:
means for searching the levels of the coding tree for a selected value from partial values of the base codewords that is less than or equal to the codeword from the bitstream with codewords;
means for skipping a fixed number of bits in the bitstream with codewords before moving to a selected level of the coding tree in response to a skip rate; and
means for calculating one of a plurality of values corresponding to the codeword based on the difference between the selected value from the partial values of the base codewords that is less than or equal to the codeword and the codeword and the index of the selected value from the partial values of the base codewords that is less than or equal to the codeword a word. 55. The device according to item 53, further further comprising means for presenting the output data to the user based at least in part on the calculated value. 56. The device according to item 53, in which the data structure contains the values represented by the base code words, and the lengths of the partial values of the base code words. 57. The device according to item 53, in which the decoding means comprises means for decoding code words to obtain values representing at least one of the video data, image data, audio data and voice data. 58. The device according to item 53, in which some of the levels of the encoding tree include codewords located in lexicographic order relative to the values represented by codewords, and each of the basic codewords is the lexicographically smallest codeword at the corresponding level in the encoding tree. 59. The device according to item 53, in which the partial values of the base code words are the removal of a fixed number of initial bits from the base code words. 60. The device according to § 59, in which, in the absence of said offset, the base codeword at a selected tree level extends, at least in part, to a remote number of start bits. 61. The device according to item 53, further containing a means of receiving code words from the encoder wirelessly. 62. A tangible computer-readable medium containing instructions for causing a processor:
refer to a data structure stored in memory, wherein the data structure contains partial values of the base codewords, the base codewords corresponding to the levels of the coding tree defining canonical codewords of variable length, the partial values being based on the shift of the base codewords by a fixed number of bits, and a skip indicator instructing the decoder to skip a fixed number of bits in the bit stream to be decoded before moving to a selected level of the encoding tree; and
decode one of the code words from the bitstream based on the partial values and the skip rate in the stored data structure. 63. The computer-readable medium of claim 62, wherein the instructions induce the processor:
search the levels of the coding tree for the selected value from the partial values of the base codewords that is less than or equal to the codeword from the bitstream with the codewords;
skip a fixed number of bits in the bitstream with codewords before moving to the selected level of the encoding tree in response to the skip rate; and
calculate one of the many values corresponding to the codeword based on the difference between the selected value from the partial values of the base codewords that is less than or equal to the codeword and the codeword and the index of the selected value from the partial values of the base codewords that is less than or equal to the codeword . 64. The computer-readable medium of claim 62, further comprising instructions for causing the processor to control the output device to present the output to the user based at least in part on the calculated value. 65. The computer-readable medium of claim 62, wherein the data structure comprises values represented by basic codewords and lengths of partial values of the basic codewords. 66. The computer-readable medium of claim 62, wherein the instructions cause the processor to decode code words to obtain values representing at least one of video data, image data, audio data, and voice data. 67. The computer-readable medium of claim 62, wherein some of the levels of the coding tree include codewords arranged in lexicographic order relative to the values represented by codewords, and each of the base codewords is the lexicographically smallest codeword at the corresponding level in the coding tree . 68. The computer-readable medium of claim 62 wherein the partial values of the base codewords are the removal of a fixed number of leading bits from the base codewords. 69. The computer-readable medium of claim 68, wherein, in the absence of said offset, a base codeword at a selected tree level extends, at least in part, to a remote number of start bits. 70. A portable wireless communications device, comprising:
a memory storing a data structure containing partial values of the base codewords, wherein the base codewords correspond to the levels of the coding tree defining variable-length codewords, the partial values being based on the shift of the base codewords by a fixed number of bits, and a skip rate instructing the decoder to skip the fixed the number of bits in the bitstream to be decoded before moving to the selected level of the encoding tree; and
a decoder that accesses a memory for decoding one of the codewords from the bitstream based on partial values and a skip rate in the stored data structure;
a receiver for receiving codewords from the encoder wirelessly; and
an output device that presents output to the user based at least in part on decoded codewords. 71. The portable wireless communications apparatus of claim 70, wherein the decoder:
searches for coding tree levels for a selected value from partial values of the base codewords, which is less than or equal to the codeword from the bitstream with codewords,
skips a fixed number of bits in the bitstream with code words before moving to the selected level of the coding tree in response to the skip rate; and
calculates one value from the set of values corresponding to the codeword based on the difference between the selected value from the partial values of the base codewords that is less than or equal to the codeword and the codeword and the index of the selected value from the partial values of the base codewords that is less than or equal to the codeword a word. 72. The portable wireless communications apparatus of claim 70, wherein the data structure comprises values represented by basic codewords and lengths of partial values of the basic codewords. 73. The portable wireless communications apparatus of claim 70, wherein the decoder includes one of a video decoder, an image decoder, and an audio decoder or voice decoder. 74. The portable wireless communications apparatus of claim 70, wherein some of the levels of the coding tree include codewords arranged in lexicographic order relative to the values represented by codewords, and each of the base codewords is the lexicographically smallest codeword at a corresponding level in coding tree. 75. The portable wireless communications apparatus of claim 70, wherein the partial values of the base codewords are the removal of a fixed number of start bits from the base codewords. 76. The portable wireless communications apparatus of claim 75, wherein, in the absence of said offset, a base codeword at a selected tree level extends, at least in part, to a remote number of start bits. 77. An integrated circuit for a wireless communication device, comprising:
a memory storing a data structure containing partial values of the base codewords, wherein the base codewords correspond to the levels of a coding tree defining canonical codewords of variable length, the partial values being based on the shift of the base codewords by a fixed number of bits, and a skip rate indicating the decoder to skip a fixed number of bits in the bitstream to be decoded before moving to a selected level of the encoding tree; and
a decoder that accesses the memory to decode one of the codewords from the bitstream based on the partial values and the skip rate in the stored data structure. 78. The integrated circuit according to clause 77, in which the decoder:
searches for coding tree levels for a selected value from partial values of the base codewords, which is less than or equal to the codeword from the bitstream with codewords,
skips a fixed number of bits in the bitstream with code words before moving to the selected level of the coding tree in response to the skip rate; and
calculates one value from the set of values corresponding to the codeword based on the difference between the selected value from the partial values of the base codewords that is less than or equal to the codeword and the codeword and the index of the selected value from the partial values of the base codewords that is less than or equal to the codeword a word. 79. The integrated circuit according to clause 77, in which the data structure contains the values represented by the base code words, and the lengths of the partial values of the base code words. 80. The integrated circuit according to claim 77, wherein the decoder includes one of a video decoder, an image decoder, and an audio decoder or voice decoder. 81. The integrated circuit of claim 77, wherein some of the levels of the coding tree include codewords arranged in lexicographic order relative to the values represented by codewords, and each of the base codewords is the lexicographically smallest codeword at the corresponding level in the coding tree . 82. The integrated circuit according to claim 77, wherein the partial values of the base codewords are the removal of a fixed number of leading bits from the base codewords. 83. The integrated circuit of claim 82, wherein, in the absence of said offset, the base codeword at a selected tree level extends, at least in part, to a remote number of start bits.

Images