DE19549491C2 - Palletised image compression with entropy encoding - Google Patents

Abstract

The method includes the steps of selecting a context model for the input symbols to be compressed, and reading the input stream of M-ary input symbols. A distribution of contexts indicated by the context model is determined, and a binary sequence is assigned to each of the M symbols in the alphabet, wherein each sequence is K bits, K being an integer greater than or equal to log2M. A reindexing table is formed, having a minimum partial bitwise entropy over a partial alphabet of symbols, for translating each M-ary symbol to its assigned binary equivalent. The input stream is translated to a bit stream which is entropy encoded into a compressed codeword stream.

Description

The invention relates to an entropy encoder according to claim 1.

Image compression is generally performed on digitized images. In order to digitize an image, the image is scanned at various locations on the image, which is often also referred to as pixels. Each pixel corresponds to a location in the image and a pixel color. For example, an image is represented by a two-dimensional arrangement of (1024 × 768) pixels, where each pixel can have a 24-bit value. Each pixel value can represent a different color, so that the image can be displayed using 2 ²⁴ or 16 777 216 possible colors.

In many applications, each pixel value becomes a specific color in the image associated with the time when the image is digitized. With the help of this assignment the color of a given pixel can be determined by the pixel value. at In other applications, the digital image is a palletized image. In a palletized picture the value of each pixel is not fixed in terms of color, but there is a pointer for an entry in a color palette table. If a palletized image is to be displayed, the color of a pixel is found by using the pixel value as an index in the Color palette table is used which shows the relationship between each pixel value contains the color associated with that pixel value.

Palletization is a selection of a few colors from the set of all possible colors, e.g. B. 256 (2 ⁸ ) out of 16 777 216 (2 ²⁴ ) colors and a method to substitute any color with one of the selected few colors. As a result, a palletized image can then be efficiently stored as a code book that (according to the example above) lists, for each of 256 indices, the complete 24-bit color description of the selected colors corresponding to the 8-bit index of each pixel in the image. This results in a space saving in which the selected colors are only completely described once and each pixel is referred to by exactly one index in the code book.

Image compression is needed for many imaging applications because of the amount of it Data important to present an image. In the above example, 24 × 1024 × 768 or 18 874 368 bits (2 359 296 bytes) needed to be a single uncompressed To display image or 8 × 1024 × 768 bits (6 291 456 bytes) in the case of a palletized image needed. Compression is even more important with one Video recording, which is a continuous image stream with a high frame rate, for example, requires 30 frames per second. Compression not only saves Storage space for image storage, but compression also allows one efficient image transmission over channels of limited bandwidth. Of course, at Using moving pictures, the pictures again quickly enough be decompressed so that they can be displayed at the frame rate. Even though the compression can be done beforehand, often requires the respective application real-time decompression. In embodiments in which the compression in ahead, the input to the decompressor is a block of compressed data, which are stored in a memory and at a time after the Compression can be decompressed.

A measure of how well compression works is the compression ratio. The Compression ratio is the ratio of the size of the uncompressed data to the data compressed data. Another measure of the goodness of compression is how data can be compressed and decompressed quickly. In many applications the compression speed is not as important as the decompression speed, since compression does not have to take place in real time.

The compression ratio is often improved the more the character of the data is known is. Normal digital images are generally the result of quantizing Waveforms; hence they retain many of the characteristics of the original signal, such as  the presence of relative continuity of pixel color values. With palletized This continuity is usually not present in images and cannot be added directly can be used to improve the compression ratio.

In the case of palletized images, the color difference between two pixel values cannot necessarily derived from the pixel values, since the pixel values are not pixel colors represent. Instead, they just put indexes in the code table or the color palette- Table of color values. The difficulty of compressing palletized images can be as the difficulty to be generalized to a sequence of symbols entered compress, with each symbol entered one symbol in one M symbols having (M-ary) alphabet.

By Y. Chin, et al. is in "Lossy Compression of Palettized Images", ICASSP-93, Vol. V, Pp. 325 to 328, an example of lossy compression of palletized Described in pictures. Although lossy compression generally increases leads to better compression ratios, the distortion in many applications not tolerated, which results from the lossy compression, and it lossless compression is consequently required. With lossless compression a signal is compressed into a compressed signal which decompresses can be used to get the original signal exactly from the compressed signal recover so that no distortion is consequently introduced.

A method is known from EP 0 562 672 A2, in which an image by means of a Data compression is shown. The image is divided into sections of a predetermined size divided in order to then distribute gray values over all picture elements of the picture determine. Then a certain predetermined scale is given in for each area Units of picture elements determined. Then each area can be divided into individual cells are divided, each for a number of points or picture elements include. Each of these points or picture elements is determined by two coordinates, the cells each having a dimension on the order of the determined dimension exhibit. Image elements are recognized in each cell and in two categories divided. Then the recognized basic elements into certain transmission elements  or sub-transmission elements converted. These are then either saved or transfer.

EP 0 574 746 A1 relates to a method for real-time compression and for real-time decompression of video signals that on the one hand convey movement and on the other others are digital. A Huffman table is used to do the coding as well Realize decoding. A fixed table must be used here so that the Sender and receiver, d. H. the encoder and decoder, exactly the same Table must be known and must also be used accordingly in order to become one useful result.

In the literature reference Rabbinani, Majid: Digital image compression techniques, SPIE 1991, S. 15 to 19, 33 to 36, 58 to 59 is presented in compression processes, in which entered symbols of an alphabet with a number of symbols, the pixels or Represent picture elements of an image, compressed to code words. These code words can be used to losslessly decode entered symbols. By means of the Bit position and a context symbol can be used to determine a context for each bit, for which a binary entropic code is determined. The latter are used to change the bits to encode the binary represented symbols.

From the foregoing, it can be seen that an improved entropic encoder is needed for pictures.

This object is achieved by an entropic encoder with the features of the claim solved. Appropriate embodiments emerge from the subclaims.

The advantages achievable according to the invention are based on a context which is characterized by Context symbols are created, which are also entered symbols, with the following Features: a symbol rearranging device with a symbol input of the Entropy encoder is connected to input symbols from an input Rearrange current order into a rearranged current order; a context Model unit connected to an output of the symbol rearranging device is which symbols are entered in the rearranged stream order and  entered symbols, if necessary, stores a context of a current symbol entered from the context symbols of a current symbol entered determine; a probability estimation unit that is connected accordingly to that current entered symbol and a context of the current entered symbol of the context model unit to get the probability of the current estimate entered symbol, which occurs in its given context, and a bit generator connected to the probability estimating unit Generate code words based on probability estimates, which are determined by the Probability estimation unit are created, the code words being a compressed one Representation of the rearranged stream of symbols entered, where the symbols rearranging facility rearranges entered symbols so that a rearranged Distance between the currently entered symbol and the context symbols of a currently entered symbol in the rearranged stream is greater than one original distance between the currently entered symbol and the context Symbols of the currently entered symbol in the entered stream.

One method of creating the reindex table is to: assign an index to an entered symbol at a specific point in time. The an index associated with an entered symbol is a function of the bits of the index, the Distribution of the entered symbols, the contexts of the entered symbols and what Symbols and indices are already assigned. A very specific, entered Symbol is assigned an index from the remaining, unallocated indexes, what creates the minimal, partial bitwise entropy that corresponds to the already assigned indices and is added to the entered symbols to which the indices are assigned. Even though this method cannot achieve the global minimum, bitwise, entropy required for one given set of symbol distributions and contexts is possible, it creates a good one "Local" minimum for a feasible set of calculations.

The re-indexed, entered data pass through a context model unit, a Probability estimation module and a bit generator, the output data of which are compressed data. The reverse process is applied to the data decompress. If necessary, the re-indexing table is overhead from the  Compressor passed to the decompressor. The reindex table is not necessary if it is determined in advance.

According to the invention, context models that use a small storage space are however, give good compression for palletized images through a contextual Model unit created. In another aspect of the invention Compression and decompression performed in parallel, and the entered values The compressor will be buffered and reordered to take up space between you and the bits that encode the context and coding of the Set coding bit. The space between a bit and its context bits are arranged accordingly so that a delay is large enough so that the context bits be fully encoded when the decompressor reaches the bit whose decoding depends on the context bits. The decompressor contains a device to do this to perform opposite orders so that the output values of the decompressor are the same as the input values to the compressor. In some embodiments run through the output values of the parallel compressors over a single one Transmission channel, whereby a small overhead is required to connect the channels to separate the various parallel decompressors.

In specific embodiments, the symbols entered are from an alphabet selected from 256 symbols; the re-indexing process is carried out accordingly programmed digital computer, the compression process is not in Real time, and the decompression process is performed in real time. at The re-indexing table occupies 256 bytes in the 256 symbols entered. In a specific In one embodiment, the re-indexing table is based on the probability distributions of symbols entered over a group of images or parts of images, and a new one The re-indexing table is sent from the compressor to the decompressor for each group of Transfer pictures or parts of pictures.

The invention is based on preferred embodiments tion forms with reference to the accompanying drawings in individual explained. Show it:

Fig. 1 is a block diagram of a data compression system including a compressor and a decompressor;

Fig. 2 is a block diagram of the compressor shown in Fig. 1;

Fig. 3 is a block diagram of the decompressors shown in Fig. 1;

Fig. 4 is a detailed illustration of a memory in which a context bin probability table is stored;

Fig. 5 is a block diagram of a data compression system, which has a Reindexiereinheit for reindexing of M symbols according to a Reindexier table;

Fig. 6 is a block diagram of an index optimization unit which generates a Reindexiertabelle, which stop on the data into data to be compressed based one block;

Fig. 7 is a flowchart of a process by means of an index optimization unit is performed to determine a geeigne te Reindexiertabelle,

Fig. 8 is an illustration of various context models that are executed by the context model units of Figs. 2 and 3;

Fig. 9 is a block diagram of a parallel compression system;

Figure 10 is a block diagram of a single channel parallel compression system;

Figure 11 is a block diagram of the operation of a reorder buffer.

Fig. 12 is a logical representation of a record showing a context state, and

13 (a) and 13 (b) Time graph of a pipeline coder which can be processed as an input symbol at a certain time Fig. More.

The description below is divided into several sections Splits. The first section describes entropy coding. The second section describes reindexing to a compress sion to improve when entropy coding is applied. The third section describes how context modeling a Entropy coding performance through more accurate estimation symbol probability opportunities can improve. Context modeling can with and oh reindexing can be used. The fourth section be writes how parallel compressors and decompressors verbes be improved to improve the data rates. Finally be the fifth section writes how buffering is applied to facilitate parallel implementations and one more Maintain causality to decompress data in parallel to be able to.

Entropy coding

Entropy encoders are in the field of data compression known. An entropy encoder, which entered symbols in out encoded code words leads to a lossless process through, d. H. the original symbols entered can be al be extracted from the output code words. On Entropy coder is so named because it tries to make a compress tion ratio that is close to the theoretical, upper The limit value is determined by the entropy of the entered sym bole is dictated. An encoder and a decoder are in space generally designed so that the decoder reverses the Codie rers is. The term "coder" denotes a device which ent is neither an encoder nor a decoder, or both.

All the details of an entropy coder are not discussed here wrote; only certain details are described, which relate to the present invention. In one Entropy encoders are input symbols from an input read and one or more entered symbols do this Output a code word. The special (s) entered Symbol (s) and code word combinations output are indicated by the  required code. The code used in one optimized system is based on the probability distribution of the entered symbols selected. In an adapti ven entropoie coder, the code used can be true change probability distribution. If the probability distribution is well estimated or well known and codes used for this probability distribution are optimized, there is good compression.

If the symbols entered are bits (i.e. a symbol ent is neither 0 or 1), the encoder is a binary entropy co the. Binary entropy encoders are compared to other entropy Coders preferred because the simplicity of the hardware is important is. An ABS encoder is a binary entropy encoder (in an am U.S. Patent Application S.N. Applicant's 08 / 016,035 are a "Procedure and Setup device for parallel decoding and encoding of data " ben. This application is above and will follow with the reference "ABS coder"). In a binary Entropy encoders are input symbols to a bit gene rator created, which one or more bits as one given value and a code word for this entered and assumes accepted bits. If the bit generator has an On show the probability distribution of the zeros (0's) and the ones (1's) in the input stream can be provided he use this information to look up the code to to use him. On the other hand, the bit generator instru what code will be used, what is by a Probability Estimation Module is set. The codes can change when the bit generator determines that the output code words too long or too short for the average neuter probability distribution of the entered symbols are. For more details, see the ABS encoder referenced application.

It is important that the probability distribution of the zeros  and ones are well appreciated. If a binary entropy co dier is used, the pixel values of one must be added compressing image, which symbols from an M-values holding alphabet, a so-called M alphabet, where M greater than two, can be put into binary form. As in turns out in some cases, the special, ge chose binary representation the entropy of the bit stream contained in the binary entropy encoder has been entered, and as above executed, entropy sets a limit on how far a bit Ge nerator can compress the input bitstream. Where the Un is not clear, a "pixel input symbol" or a "pixel symbol" is used to refer to an element of the compri reference data before it is converted into binary form what is then an input value to a compressor, and a "binary input symbol" is used to point to a given symbol to refer to, which is in a bit generator of a binary entropy encoder is entered.

In the case of non-palletized image data, the pixel symbols can be brought into binary form in accordance with color information, which is included in the pixel value. For example, a pixel value should assume one of 256 values and the accepted value should indicate the color of the pixel. In this case, if similar colors are represented by similar 8 bit values, it can safely be assumed that 00001100 and 00001101 are more similar in color than 00001100 and 11111111. In a number of examples discussed here, M is 256, and symbols in the so-called M Alphabets are represented by K-bit binary representations, where K = 8. Of course, other values of M and K are also possible, including values of M that are not exactly powers of two, the only limitation being that K ≧ log ₂ M.

Because in the example in the typical picture neighboring pixels often may have similar colors, the number of bits required are, to express each pixel, are reduced,  that each pixel by the difference in the pixel color value between the pixel and a neighboring pixel is displayed. This preliminary Data processing generally leads to a better comm pression, since the number of bits needed to add values is related to the entropy of the group of values. However, in the case of pixels it runs in a palettier th picture a very specific relationship between two colors not a correlation between the bits of such 8 bits Representations beyond. For example, in a pallet picture 00001100 and 00001101 can be very different colors, and 00001100 and 11111111 could be almost identical colors his. It all depends on the pallet.

As a result, the compression of palletized images should be reduced can be improved by the probability distribution of the bit current is estimated accurately and by the binary representation of the entered pixel symbols is chosen, which is a low Entropy bitstream on a bit generator creates. Solutions of the First problems are explained in the section below "Context modeling" are overwritten, and solutions of the The last mentioned problem are discussed in the section that is overwritten with "reindexing".

decompression

So that a compressor entered symbols from output code words decoded correctly, the decompressor must know which one Code for which code words is used, if not the code rer is an adaptive encoder and a fixed code used. Since the overhead of broadcast data indicating the code would reduce the usefulness of the compression, the Decompressor to be able to display the codes by the code words have been used to extract. This is possible if the code is at a given binary Input symbol has been used by a Probable is estimated, which only depends on bits that are in Codewords have already been encoded. If this is correct  can the decompressor using a feedback loop and stores information about the entered symbols if they are recovered, specifying which code is only data is used, which are already decoded.

Context Modeling

The more accurate the probability estimates are, the more are the codes used regarding the compressed data op timiert. One way to miss the probability estimates nern, is to estimate probability for more to maintain subject contexts. The context is then one entered pixels the pixel or pixels in a defined Relationship to the entered pixel. For example, in a context model the context of an input pixel determines the pixel above the entered pixel. If the context of a bit in the bit stream input to a bit generator the context is determined by pixels and / or bits in established relationship to the input bit fen. The relationship is determined by the context model.

Figure 8 is a graphical representation of five context models. In each model, a pointer P indicates the bit position of the current input bit in the bit generator of the current input symbol to which the context matches. In all contexts shown, the context is derived from any subset of the bits of the current input symbol, the bit position indicated by P, and the bits of the symbol for the pixel in the image above the pixel for the current input symbol, which is called the symbol "above" is referred to. Of course, other pixels, such as the pixel to the left of the current pixel or other context models, not shown, can also be used as well.

Each context model specifies how the context is determined is, and the number of possible contexts can easily  be calculated. For each context model there is a probability probability estimate, and this probability Estimation is used when a symbol has this context occurs. The term "contextual subject" is often used to refer to show the context in which a symbol falls.

In Fig. 8, the context model (a) is a bit position context model. In this model, the only context is the bit position, which requires K context subjects. With K = 8, this is a reasonable number of context subjects.

The context model (b) is a bit level context model, wel ches is so named because it is a set of M contexts for every bit level there. Hence there are (M × K) contexts (2048, if M = 256 and K = 8). A mark that is special To be mentioned is that the contexts of the current Pixels in models (a) and (b) are independent of the Pixels used to represent the current pixel len.

The context model (c) is a bit-dependent context model, the context being determined by the bit position when the current bit and the previous bits are in the current symbol. In this model there are ( ^2K - 1) contexts (one (1) context when pointer P is on the first bit, two (2) contexts on the next bit, etc.). For K = 8, this model creates 255 contexts. Note that the context of a bit cannot be relied on the values for bits that are not yet encoded, so the decoder needs undecoded bits to determine how to decode a bit. (A processing sequence from top to bottom and left to right is assumed in Fig. 8).

The context model (d) is a good context model, which all bits of the pixel "above" and the bits of the entered Pixels taken into account which are before the current bit.  

This context model links the information of models (b) and (c) and consequently creates M × (2 ^K - 1) contexts. Although this context model could create a fine division of the contexts, the memory necessary to store values for these contexts can prevent a hardware implementation of a compressor or decompressor. For M = 256 and K = 8 this model requires 65 280 contexts. If each context requires 8 memory bits, this context model would require 64K memory for that purpose. However, an advantage of this contextual model over models (a) and (b) is that the performance of a compressor using this model is good regardless of the digitization used, since older bits of the current bit are used to To determine contexts.

The context model (s) is a mixed context model, which is to be seen as a good compromise between memory requirements and the quality of the context model. In this model, the context is the (P - 1) previous bits of the current input symbol and the (K - (P - 1)) last bits of the symbol above it. This requires 2 ^K contexts per bit position (2048 contexts for K = 8). The context model is generally established between the compressor and the decompressor. However, the compressor and the decompressor cannot be prevented from switching context models as long as they agree which context model is used at a given time. If changeable context models are permitted, a mechanism must be in place to ensure that the context models are the same, for example a communication channel overhead that indicates the current context model.

The above discussion of entropy coding and context modeling is provided as a basis for explaining the compression system shown in the figures, which will now be described. In Fig. 1 a is off guide form a data compression system 100 according to the invention is shown, to compress data comprising symbols in a so-called M-alphabet, and the execution form includes a compressor 102, a channel 104 and a De compressor 106. The input values to compressor 102 are a stream of input pixel symbols, each representing a pixel value in an image, although other types of data could also be used.

The entered pixel symbols are converted by means of the compressor 102 into compressed code words which have been transmitted to the decompressor 106 via the channel 104 . Decompressor 106 retrieves the original input symbols or their equivalent only from the code words received and possibly from knowledge of the operation of compressor 102 and a re-indexing table (if used). In some cases, the original input symbols are translated from a representation to a re-indexed representation, and the data compression system treats the decompressed symbols without actually reversing the re-indexing.

An advantage of the invention lies in the increased compression ratios for input symbols, which are pixel values in a palletized image. Therefore, the compression created by the invention can be used whenever the channel 104 has a limited bandwidth. Examples of the channel 104 have a communication channel which is provided between an image memory and a display unit, or a storage unit and a device for a stored image. Channel 104 may also represent the transmission of compressed data in the form of devices with memory, such as game cartridges, which data is compressed at one location and transmitted to another where it is decompressed. In a later state, the bandwidth limitation of the channel 104 leads to a limitation in the storage capacity which is available in the storage unit.

In Fig. 2, compressor 102 includes a text model unit 108 ( 1 ), a probability estimation module (PEM) 110 ( 1 ), a bit generator 112 ( 1 ), a context memory 114 ( 1 ) and a context subject probability memory 116 ( 1 ). The numbers in parentheses are used to distinguish like numbered elements found in both compressor 102 and decompressor 106 . Reference numerals of elements which are generally usable in the compressor 102 or in the decompressor 106 are shown without the numbers set in parentheses. Although the connection is described using a particular compressor 102 shown in FIG. 2, other compressors may work just as well. For example, not all compressors require the coupling shown between the PE module 110 and the bit generator 112 , and a PE module 110 may rely on data other than that from the bit generator 112 to update its likelihood estimates.

The context model unit 108 (FIG. 1 ) is connected to the input of the compressor 102 and thereby receives the entered pixel symbols. If a re-indexing unit is used, the input to compressor 102 is a stream of re-indexed, given pixel symbols. A context model unit 108 ( 1 ) has two outputs, a context subject output and a decision output, which are both inputs to the PE module 110 ( 1 ). If the input is represented in binary form in a symbol of K bits, then the context model unit 108 ( 1 ) makes K binary "decisions" per input pixel symbol. The context model unit 108 ( 1 ) also issues a context subject identifier for each decision. A context model unit 108 ( 1 ) is also connected to the memory 114 ( 1 ) to address data elements therein and to read / write data regarding the occurrence of symbols in the input stream and, if necessary, to To determine the context of the entered pixel symbol, which is represented as its input. The particular layout of data elements in memory 114 ( 1 ) depends on the context model.

The PE module 110 ( 1 ) receives the two input values from the context model unit 108 ( 1 ) and has two outputs to the bit generator 112 ( 1 ): a probability class (Pclass) output and a result output. The PE module 110 ( 1 ) is also connected to the memory 116 ( 1 ) to address data elements therein and to read and modify values of these data elements, and is also connected to the bit generator 112 ( 1 ) to to get a table update signal. The PE module 110 (1) for each decision input at PE module 110 (1) outputs a result bit and a PCLASS value. The result bit depends on the decision input bit and a most likely bit symbol (MPS) for the context bin associated with the decision input bit. The result bit indicates whether the decision input bit is the MPS symbol or not. The opposite of the MPS symbol (which is either 0 or 1 because the bit generator encodes binary symbols) is a least likely symbol (LPS).

Bit generator 112 ( 1 ) receives a stream of result bits and their associated Pclass values. Thus, the bit generator 112 ( 1 ) receives K result bits and an associated Pclass value for each pixel symbol that has been input to the context model unit 108 ( 1 ). From these two input streams, the bit generator 112 ( 1 ) generates a stream of compressed code words, as explained in the ABS encoder mentioned at the beginning. With a properly sized bit generator, the result bit stream can be recovered from the output code words, and the code words have fewer bits than the input bit stream. The Pclass value is used to select a code to use to encode the bit stream. On the other hand, the PM module ( 1 ) can store the probability distribution of the MPS and LPS symbols for each context and create the probability on the bit generator.

As stated in the application relating to the ABS coder, are some binary codes at certain probabilities of symbols (0's and 1's) optimized in the input stream. For example, if 0's and 1's at the input to a bit generator are equally likely, then an optimal one Code one without compression, d. H. a bit that's turned on for each given bit is given. The optimization of a code is there determined by how close the average code word length approaches the theoretical entropy of the input bits.

Since the compressor 102 acts on a pixel symbol at a certain point in time and the PE module 110 and the bit generator 112 act on a bit of this pixel at a certain point in time, the terms "current" pixel and a "current" bit have a corresponding meaning and are used in the following description to explain the operation of the compressor 102 and decompressor 106 . Of course, symbols cannot be acted on at a certain time in a complicated compressor; however, the term "current" symbol can still be used. Of course, elements can also be operated in a pipeline type so that the current symbol for one stage differs from the current symbol for another stage.

In a compression operation, the compressor 102 sequentially picks up each input pixel symbol and outputs a stream of code words, although a code word is not necessarily given for a predetermined number of input symbols. When processing a pixel symbol, the symbol is expressed in binary notation and the binary representation is entered into the context model unit 108 ( 1 ). The context model unit 108 ( 1 ) evaluates the binary representation on a bit basis, with each bit being delivered to the decision output of the model unit along with a context bin ID that identifies the context of the bit. If, for example, the context model used is the bit position context model, the context compartment ID only shows the bit position in the binary representation of the entered pixel symbol. For the bit position context, only one memory must be in the current bit, but for different context models, each pixel that affects the context of another pixel is stored in memory 114 . If the context model contains the pixel above or above the current pixel, it would be sufficient to store one line of pixels in memory 114 .

If the input symbols represent pixel colors in a two-dimensional array of pixels that form an image, the scanning order is not so important for a text block, context being generally provided by the letters near a current letter. However, in this discussion it is only assumed that the two-dimensional arrangement is keyed into the compressor 102 in a predetermined order, for example from an upper to a lower line and from left to right in each line. With this scan order, the pixel above the current pixel can be used for context since the pixel above it has already been encoded and it is necessary at this point to create a context for decoding the current pixel, and since in a common picture the pixel directly above the current pixel is often a color that is likely to match the current pixel.

This is nothing special about the context pixel that the Pixel is above the current pixel, and although many context Mo dents are possible requires successful, lossless Decompression of the compressed data stream that the ak current bit context model does not contain bits or symbols, wel have not been decoded at the time the  current bit is to be decoded. Some examples are given to this Clarify point.

If the context model is a "bit position" context, the is Context of the current bit the bit position in the current entered pixels, and the context ID for that context would represent the current context, which in the case of K Bit bi symbols depicted here K creates contexts. However, if the current bit context determines the code that has been used to encode this bit, the context must be known before the bit can be decoded correctly. Hence the Processing order ensure that the context of the ak actual bits is known before the bit is decoded.

For each bit that has been delivered by the context model unit to its decision output, the PE module 110 ( 1 ) accepts the decision and the context subject ID for this decision. Based on this decision regarding the MPS symbol for the context subject of the decision, the PE module 110 ( 1 ) outputs a result bit and an assigned Pclass value. How the MPS symbol and the Pclass value are determined is explained below.

The Pclass value is an estimate of how likely the MPS symbol is in the bit stream that has been provided by the context model unit 108 ( 1 ) to the decision signal line. Although the probability distribution of the MPS and LPS symbols in the input data can be determined precisely by sampling all the data entered, the compression should not depend on the full probability distribution, since it is not available on the decompressor before the code words are decoded , Therefore, the estimate, which only depends on previously coded bits, is used. As explained in the application relating to the ABS encoder, the bit generator 112 ( 1 ) uses the Pclass value to determine which code to use to encode the result bit stream. Because code is optimized over a range of probability distributions, a Pclass value defines a range of probability values, hence the named probability class. These probability class values are stored in memory 116 , one per context. Memory 116 also stores the MPS symbol for each context. Fig. 4 shows the layout of the memory 116th Initially, the Pclass value for each context is a Pclass value indicating that the probability of the MPS symbol is 50% (and that of the LPS symbol is 50%) and the MPS symbol can be chosen from 0 and 1 ,

As stated in the application relating to the ABS coder, be is a method of compressing data into an entropy encoder in using code that runs long of the MPS symbol to a code word which reduces the length of the run. The original stream is by Abge ben reconstructed a number of MPS bits, which by the Code word have been displayed, followed by the LPS symbol, au If a pass of MPS bits that is longer than a maxima The pass is divided into more than one word.

After a code word is given, the bit generator 112 ( 1 ) signals the PE module 110 ( 1 ) whether the maximum throughput of MPS bits is present or not. This display is provided on the table update signal line. When the maximum pass length has arrived, bit generator 112 ( 1 ) signals PE module 110 ( 1 ) to update the Pclass table to indicate that the MPS symbol is more likely than the previously estimated one. If the maximum run length has not been reached, the Pclass value is changed to indicate that the MPS symbol is less likely than the previously estimated one.

Decompression is exactly the opposite process to compression: code words are decoded into the result bit stream, the delivered bit stream is converted into the decision bit stream and compiled into pixel symbols, which are then delivered. In Fig. 3, the decompressor 106 includes a context-model unit 108 (2), a PE module 110 (2), a bit generator 112 (2), memory 114 (2) and 116 (2). The interconnection of the modules in the decompressor 106 is the inversion (interconnection) of the compressor 102 for the bit stream, but the same direction applies to the information in the bit stream, so that two feedback loops are formed.

A feedback loop determines the code used to encode a code word. Knowing the codes used, bit generator 112 ( 2 ) can decode the bit stream. The codes used are determined by the Pclass value created by the PE module 110 ( 2 ). The Pclass value, as with the PE module 110 ( 1 ), is determined by the context and signals on the table update line. The context is determined from information about previous bits and pixels stored in memory 114 ( 2 ). The feedback loop has a certain delay, which can be measured in a number of bits. The delay may change depending on the code words received, and in most embodiments, the delay may be a maximum. If the delay maximum is X bits, then the current bit context should not depend on the X bits that precede the current bits in the decision bit stream. This ensures that the current bit can be decoded from information that has already been processed by decompressor 106 . This requirement brings with it special problems which the invention solves, the data having to be decompressed in parallel.

reindexing

As stated above, binary entropy encoder data can compress up to a theoretical limit. A binary representation or digitization of palletized data can be used to display an image as conveniently as any other (by changing the color palette table); however, it turns out that some binary representations or digitizations create less entropy for some contextual models than others. This advantage can be created by adding a re-indexing unit at the front end of the compressor 102 , the re-indexing unit selecting a binary representation which creates a low entropy for the data to be compressed. The reversal of a re-indexing unit, a de-indexing unit, can be accommodated at the output of the decompressor 106 or the definition of the color palette can be modified to account for the change in alphabetical space as a result of the re-indexing. In some embodiments, the re-indexing table can even be set before compression and stored in both the compressor and the decompressor.

FIG. 5 shows a re-indexing data compression system 158 which has a re-indexing unit 160 connected to the input of the compressor 102 , a re-indexing table 162 which is a (2 × M) table and optionally a de-indexing unit 164 . A signal line 166 is shown to indicate the transfer of the re-index table to decompressor 106 when re-indexing is not fixed. Since the pure indexing table 162 is only (2 × M), no noticeable overhead is brought about by the transmission together with the compressed data.

The operations of the re-indexing unit 160 and de-indexing unit 164 are straightforward and are specified by a special re-xing table 162 . The reindexing unit 160 indexes into the table 162 with the aid of a pixel symbol from the input thereof and reads out the corresponding reindexing value. This value is output as the binary symbol input to the context compressor 102 . The reindex values in the table 162 are such that the operation is reversible, ie exactly one entry in the reindex table 162 applies to each predefined reindex value. Thus, if the output of decompressor 106 , which is the same as the compression system delay, as the input to compressor 102 , is applied to de-indexing unit 164 , if used, de-indexing unit 164 performs the opposite transformation and outputs the index , which contains the original M-pixel symbol. As mentioned above, if the actual original M symbol is not needed and a known binary representation can be used, the de-indexer 164 is not needed.

FIG. 6 is a block diagram of a clean text table generator 200 used to generate the reindex values in table 162 from an input data block. The generator 200 can be implemented either in dedicated hardware or in a suitably programmed digital computer. Using a digital universal calculator is often cheaper, especially for one-time compressions, although it is slower than dedicated hardware. Most applications may allow compression to be done beforehand rather than in real time.

The input to generator 200 is a block of pixel symbols. This block may be the entirety of the images to be compressed, a scanned subset of the images, or may even be parts of images. However, if compression is not to be performed in real time, there is generally no problem in using the entire data set as the input to the generator 200 . In some embodiments, the image data is arranged in blocks of 16 or 256 images and a re-index table is created for each block of images.

In FIG. 6 200 includes the generator comprises the sequence festle constricting module 204, a conditional Wahrscheinlichkeitsvertei lungs accumulator 206, a conditional probability table 208 for holding values P _c (S _i) for (i: 0 to M - 1 and c: 0 to N-1), where P _c (S _i ) is the probability that an input symbol S _{i occurs} in a context C _c , a context probability table 209 for holding values P (C _c ) for (c : 0 to N - 1), where P (C _c ) is the probability that the context C _{c occurs} in the input data block, a greedy entropy minimization unit, the output of which is connected to the reindex table, a partial conditional Probability table le 212 ("the PCP table") for holding values PCP (c) for (c: 0 to N-1), where PCP (c) is the partial conditional probability of the context C _c , as explained below, a partial position probability table 212 ("the PPCP table") for holding Values PPCP (k, c) for (k: 1 to K and c: 0 to N-1), where PPCP (c) is the partial conditional position probability of the context C _c and the bit position k, as explained below , and a partial bit-wise entropy table 215 for holding values PBE (j) for (j: 1 to M-1), as will be explained later. Using the various tables will become clearer after reading the description of the flow chart of FIG. 7.

The generator 200 operates as follows. The order-determining module 204 converts each symbol from the data block 202 into an ordered symbol S _i , where i is 0 to M-1. Since the effect of the re-indexing table 102 is to reorder the symbols, the particular order used is of less importance than the operation of the rest of the system. In one possible embodiment, symbols in data blocks 202 are already stored as binary K bits (K ≧ log ₂ M) values, in which case module 204 is not required. In one embodiment, S _{0 is associated with} the most common pixel symbol in data block 202 , S _{i with} the next most frequent and so on until S _{M-1 is associated} with the least common pixel symbol. This is sometimes known as the 0th order frequency distribution or as a contextless symbol frequency distribution.

The module 204 , when used, reads pixel symbols from a data block 202 and passes the binary representation of the ordered symbols to the accumulator 206 , which collects probabilities in tables 208 and 209 . The entropy minimization unit 210 then reads the contents of tables 208 and 209 , reads values and writes values in tables 212 , 214 and 216 to produce their main output value, which becomes the contents of table 162 . In some embodiments, the minimizer 210 includes an internal table to which indexes from table 162 are associated, and in other embodiments, the minimizer 210 determines which indexes are associated when looking into table 162 .

The operation of the reindex table generator 200 can best be understood from the flow chart shown in FIG. 7 in conjunction with FIG. 6. The flow diagram shows blocks 300 through 328 . Block 300 indicates the start of a process to generate a reindex table from a data block. The steps in blocks 302 and 306 are usually carried out by the accumulator 206 , the step in block 304 by the module 204 , the rest being usually carried out by the entropy minimizing unit 210 . The process flow goes in block number order from block 302 to block 328 and either ends after block 328 or goes back to block 316 as shown below.

In block 302 , accumulator 206 deletes tables 208 and 209 . In block 304 , module 204 , if used, sorts the symbols entered into S ₀ through S _M-1 , and in block 304 , accumulator 206 fills tables 208 and 209 . When filled, the table 208 contains an entry for each combination of a context and an entered symbol, an entry (c, i) of the table 208 indicating the probability of the symbol S _i that occurs in the context C _c . When filled, the table 209 contains an entry for each context, an entry c of the table 209 indicating the probability of the context C _c , which occurs regardless of the symbol. There are M possible values for i, and the contexts are {C _0,. , ., C _n-1 }, where N possible contexts occur. As a result, table 208 contains (M × N) entries and table 209 contains M entries.

In one embodiment, the context is the symbol for the pixel above the pixel from which S _{1 was} taken, in which case N = M. In other embodiments, the context is the symbol for the pixel on the left (N = M) or the context is two symbols (N = M ² ). However, a two symbol context would require M ² contexts for each symbol entered, and thus tables 208 and 209 would be very large for most values of M. In still other embodiments, there are fewer than M contexts. For example, if all possible symbols are grouped into less than M groups and the context for a context pixel is the group that receives the pixel above the context pixel, then fewer than M contexts will occur. For example, if all symbols were grouped into groups of four symbols, NM / 4 would be. For this reduction in context to be useful, an intelligent device for grouping symbols should be used.

In block 308 , the minimizer 210 sets a loop variable i to zero and initializes the reindex table 162 , the PCP (c) table 212 and the PPCP (k, c) table 214 . Although block 308 is shown after block 306 , its order could be reversed since neither step depends on the other's result. The reindex table 162 is initialized to use flags or other known means to show that all entries in the table 162 are empty and no indexes are associated. The PCP table is initialized by setting PCP (c) equal to 0 for all c, and the PPCP table is initialized by setting PPCP (k, c) = 0 for all k and c. In block 310 , the minimization unit 210 assigns an index 0 to S ₀ and updates the reindex table accordingly. In block 312 , the minimizer 210 updates the PCP (c) table for the added index S ₀ by setting PCP (c) = Pc (S ₀ ) for each c. In block 314 , the minimizer 210 updates the PPCP (c) table for the added index S ₀ by setting PPCP (k, c) = Pc (S ₀ ) for each (c, k) pair in which the k -th bit of the base 2 representation of the assigned index "1". This step is not necessary if the first associated index is 0 because none of the bits of the index is 1; however, this step is required if a non-zero index is assigned first.

In block 316, the minimization unit 210 increments i, which results in i = 1 on the first pass through block 316 . The process in blocks 316 through 328 is repeated for each i between 0 and M-1, with an index in the reindex table assigned to a symbol in each loop. Of course, the last loop, in which i = M - 1, is actually not necessary, since only one index remains at this point to be assigned to S _M-1 .

In block 318 , the PCP table for the symbol S _{i is} updated accordingly to PCP (c) = PCP (c) + P _c (S _i ). After this step, the probability, which is determined by the context C _c and the alphabet of symbols A '= {S ₀ ,. , , S _i } is given that the received symbol in the alphabet A 'is PCP (c), the c-th entry of the PCP table. At block 320 , a partial bitwise entropy (PBE) value is generated for each unassigned index. The unassigned indices (the number of which is M-1 at this point) are found in table 162 or are noted internally by the minimizing unit 210 . The partial bit-wise entropy for an index j and a symbol S _i , PBE (j, S _i ), which is given by the fact that the indices 0 to i - 1 have already been assigned, is given in Eqs. 1 to 3 expressed.

where A 'is the alphabet of the symbols already assigned, H (P) is the entropy of the probability distribution P, b _k (x) is the value of the kth bit of a base 2 representation of x and r (s) is the reindexing value for the symbol is s.

In block 322 , the minimization unit 210 selects the minimum value PBE (j, S _i ) from all calculated values and sets a variable X equal to the j of this minimum value. In block 324 , the minimization unit assigns 210 × the reindex value for S _i and shows the assignment in table 162 .

In block 326 , the minimization unit 210 updates the PPCP (c, k) table for S _i and X corresponding to PPCP (c, k) = PPCP (c, k) + P _c (S _i ) for all c and k, where: b _k (X) = 1. After this step, the probability is determined by the context C _c and the alpha bet A '= {S ₀ ,. , , S _i } is given that the kth bit of a re-index value for a received symbol is 1 and the received symbol in the alphabet A 'is PPCP (c, k). In block 328 , the minimizer 210 checks to see if two or more reindex values have not been assigned. This is the case if i is M - 2 or less in this block, which means that (M - 1) indices have been assigned. The last index is assigned to S _M-1 .

The denominator in Eq. (2) is PCP (2), which is updated in block 318 by adding p _c (S _i ) each time a symbol S _{i is added} in the alphabet A '. The counter in Eq. (2) is PPCP (k, c) as it would be if j were the index for S _i . Since the updated value of PPCP (k, c) depends on the particular mapping from index j to S _i , PPCP (k, c) is updated in block 326 after the selection of X = j. Of course, since H is a bitwise entropy function, Eq. (2) to be calculated only for either β = 0 or 1, since: P (0) + P (1) = 1. Since H is also a bitwise entropy, Eq. (3) actually to:

H (x) = -x.log ₂ x - (1 - x) .log ₂ (1 - x).

The end result of the process illustrated in Figure 7 is the mapping of re-index values in re-index table 162 , which is used for re-indexing unit 160 to re-index entered pixel symbols into a binary representation, which generally improves Compression leads as if the reindexing was not carried out.

Parallel processing

The previous sections describe how compresses conditions through the use of adaptive coding, Context models and reindexing need to be improved. In one Decompressors require adaptive coding and contextual Model feedback loops because of the way on which a bit is decoded from a code word stream, from In formation depends on previous decoded values and symbols. If the decompressor requires it, the code word stream in real Decoding time quickly could involve parallel processing can be used to create a larger data flow.

Such a parallel processing implementation is de. With the help of the bit position context model (see FIG. 8 (a)), each bit in a pixel value is processed by one of eight parallel coders, each bit position being left to each coder. With this arrangement, the context of the bit in a coder is known because the bit position is the only context, and the bit position is known to the coder because it is a constant for a specific coder. However, if other context models are used to provide better compression, connecting a encoder in parallel is not that easy.

As an example, the bit-dependent context model is shown, which is shown in Fig. 8 (c). If each bit of a pixel is sent to a separate encoder, the bits cannot be decoded in parallel since bits to the left of a bit to be decoded have to be decoded first to create the necessary context for the bit to be decoded. Consequently, the use of some context models creates constraints on parallel processing. In one embodiment of the invention, a reordering buffer is therefore provided in order to enable the use of a preferred context model and also to allow parallel processing in a coder which uses these context models. Figures 9 and 10 show how a reorder buffer and its reverse, a deordering buffer, are connected to parallel coders; Fig. 11 shows the operation of the reorder buffer in detail.

In a preferred context model, two restrictions are used in a parallel encoder. If each stage has to be independent, then in the set of bits processed in a parallel cycle, none of the bits should affect the context of the other bits. Furthermore, if the context data is to be updated after each bit is processed, two bits in the set of bits processed in parallel should not share a context. Figure 12 illustrates a typical context record that contains context data that can be updated when a bit is processed with a given context. This record may be a row in a table, for example in the memory 116 shown in FIG. 4.

In Fig. 9 a parallel compression system of the invention with a Umodnungspuffer 402, a demultiplexer 404, N parallel stages 406,400 according to (0, N -... 1), a multiplexer 408 and an order-cancellation buffer 410 shown. System 400 can be used with and without reindexing. Each stage 406 has a compressor 412 , which speaks to the compressor 102 , a channel 414 and a decompressor 416 , which corresponds to the decompressor 106 .

Fig. 10 shows an alternative configuration 450 in which an interleaving unit 452 is used to link the operation of channels 414 (0,..., N-1) to a single channel 454 and a deinterleaving unit (de-interleaver ) 456 , which is used to split the operation of the single channel 454 back into the separate channels for each stage.

In FIG. 9, the symbol stream that has been input to system 400 is buffered by reorder buffer 402 . Reorder buffer 404 receives a bit stream of K bits (K = 8 in FIG. 8) per symbol, which will be explained in more detail below, and delivers the bits to demultiplexer 404 , which sends the reordered bits to the parallel encoders distributed. In the "i" th parallel stage 406 (i), the compressor 412 (i) compresses the bits associated with the stage 406 (i) into code words which in turn pass through the channel 414 (i) and are decompressed by the decompressor 416 (i) become. The outputs of all stages are then merged by multiplexer 408 and their mating is provided to de-ordering buffer 410 which does the opposite of reorder buffer 404 . Although a single context is assigned to each level in the parallel coder shown, rearrangement allows flexibility to divide the parallel levels in a different way according to the needs of the application. For example, each level could be assigned a Pclass value, where a group has a number of contexts, or a division that is specifically designed even for operation over the different levels.

Fig. 11 (a) shows a reordering buffer 502 which is used in a parallel coder with four parallel stages and the bitwise dependent context model, although it can be used in other configurations, such as 8 or N parallel stages, one different number of bits per pixel or another context model. Four input bytes (pixels) are shown on the left to be input into the buffer 502 ) and eight 4-bit blocks are output from the buffer 502 . The delivered blocks, one per parallel coding period, are delivered to a parallel encoder; consequently, each of the output blocks represents a group of bits processed in parallel. If the input bits were not rearranged, the blocks that were input to the parallel encoder could just be the four most significant bits of a pixel A (A7 to A4) followed by the four least significant bits of pixel A (A3 to A0), then the four most significant bits of pixel B (B7 to B4), etc. As can be seen from the above discussion, without reordering a set of bits processed in parallel would not be independent if the bitwise dependent model is used, since A4 has its context from a position (4) and the values of bits A7, A4 and A5 receives which are in the same group of bits.

To solve this difficulty, the buffer 502 rearranges the given bits, as shown, to the blocks (sets of bits processed in parallel) D7-C7-B7-A7, D6-C6 for pixels A, B, C and D. -B6-A6, etc. With this reordering, no bit relies on any other bit in its context group. However, if two bits still share a context, for example, D7 and C7 would use a context in the bit-dependent context model since the bit-dependent context is the bit position and the bits on the left. For D7 and C7, the bit positions are the same and there are no bits to the left of it (see above in Fig. 11).

This second limitation is solved together with the first limitation by the reorder buffer shown in Fig. 11 (b). In this buffer, pixels grouped in blocks are taken from different pixels, but also from different bit positions. If a bit position is part of the context, bits with different bit positions have different contexts. Thus, the blocks (sets of bits processed in parallel) output from the buffer 504 each have D7-C6-B5-A4, D6-C5-B4-A3, etc. up to the ninth block H7-G6-F5-E4, which are not depend on other bits in its block for context and not share a context with any other bits in its block. Entered pixels E, F, G and H are not shown, but follow pixel D.

In Fig. 11 (b), there are no pixels preceding pixel A to explain the state of the buffer that follows a reset. In the first parallel step after resetting the buffer, a block with only one bit (from a pixel A) is processed. In the next parallel step, a bit is processed from a pixel A and a bit from a pixel B. Although the staggering of bits to avoid context conflicts means that not all parallel stages are used in several first blocks after a reset, this underuse is meaningless for a reasonable image size. In the case of Fig. 11 (b), this only occurs in the first three blocks.

Fig. 13 (a) and (b) illustrate a limitation of a context-Mo model which is arranged in pipeline coder, and corresponds to the restrictions which Be, which have occurred in parallel coders. Fig. 13 (a) and 13 (b) show the progress of a pipe line coder, which the pixel A shown processed to D in Fig. 11, Fig. 13 (a) illustrates a pipeline, the bits in a non- reordered, while Fig. 13 (b) shows a pipeline that processes bits in a reordered order. In both figures, each column is overwritten with a time t and shows the bit that is processed by a pipeline stage. The stages shown in this example are one input stage, three intermediate stages (IS-1 to IS-3) and two output stages.

In this example, we want to process a bit that is available to determine the context of other bits in the first output stage, but not earlier. Furthermore, the processing of IS-1 should depend on the context of the bit to be processed, and the bit-dependent context model should be used. In the context model depending on the bit, the context of the circled bit A2 depends on A3 to A7. In Fig. 13 (a) this would mean that the processing of bits A2 in IS-1 cannot continue at least until bits A3 and A4 are all on the way to the first output stage. (Bits A5 to A7 are already beyond the first output stage. This delay would negate the advantages of ipeline processing.

To solve this difficulty, a reordering buffer such as buffer 502 can be used for the effect shown in Fig. 13 (b). In this example, bit B6 is circled, which gets its context from bit B7 which was given to the first output stage in a previous section, and is thus available when B6 is to be processed in IS-1.

Claims (3)

1. Entropy encoder, in which input symbols are encoded, based on a context created by context symbols, which are also input symbols, with the following features:
A symbol rearranging device connected to a symbol input of the entropy encoder to rearrange input symbols from an input current sequence into a rearranged current sequence;
a context model unit connected to an output of the symbol rearranging device, which receives input symbols in the rearranged stream order and stores input symbols, if necessary, to determine a context of a current input symbol from the context symbols of a current input symbol ;
a likelihood estimator connected appropriately to obtain the current input symbol and a context of the current input symbol from the context model unit to estimate the probability of the current input symbol that occurs in its given context, and
a bit generator connected to the probability estimator to generate code words based on probability estimates created by the probability estimator, the code words being a compressed representation of the rearranged stream of input symbols,
wherein the symbol rearranging means rearranges input symbols such that a rearranged distance between the current input symbol and the context symbols of a current input symbol in the rearranged stream is greater than an original distance between the current input symbol and the context symbols of the current one entered symbol in the entered stream. 2. Device according to claim 1, with a decoder, which comprises:
a second context model unit that stores decoded input data if necessary to determine a context of a current input symbol;
a second probability estimation unit, which is switched accordingly to obtain a context of the currently input symbol from the second context model unit;
a second bit generator, which is connected accordingly to receive the code words which have been output by the bit generator, in order to generate a symbol stream from the code words based on probability estimates, which are generated by the second probability estimation unit are created;
a symbol rearranging device connected to a symbol output of the second context model unit for rearranging decoded input symbols from the rearranged stream order into the input stream order, thereby causing a delay between the context symbols to be decoded and the current ones decoding entered symbols. 3. Device according to claim 2, with a device to de parallel the code words encode, the larger delay being sufficient to cause parallel decoding possible context symbols, if necessary, created by the decoder become.