JP3108404B2 - Data compression device and data decompression device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a data compression device and a data decompression device for encoding data such as a character code and an image in a byte unit or a word unit. TECHNICAL FIELD The present invention relates to a data compression device and a data decompression device having improved processing performance in image processing.

[0002]

2. Description of the Related Art In recent years, various types of data such as character codes, vector information, and images have been handled by computers, and the amount of data handled has rapidly increased. When dealing with a large amount of data, by compressing the amount of data by omitting redundant portions in the data, it becomes possible to reduce the storage capacity or to transmit the data to a remote place.

[0003] The field of the present invention is not limited to character code compression but can be applied to various types of data. In the following, one word unit of data is called a character, following the name used in information theory. Data in which an arbitrary number of words are connected is called a character string. Universal encoding is a method that can compress various data with one algorithm. There are various methods for universal encoding,
There are two typical methods. Dictionary coding and probability statistical coding.

[0004] Dictionary coding or Lem
pel-Ziv coding (name of the inventor)) Representative method: LZW method, LZSS method, etc. Stochastic coding (Statistical coding) Representative method: Multi-valued arithmetic coding, dynamic Huffman coding, etc. The appearing character string (character string length is variable) is registered in a table called a dictionary, and the next appearing character string is encoded with the position information (code length is constant) in the dictionary of the longest character string registered in the dictionary. Become Therefore, it is based on VF (Variable-Fixed coding) encoding in which a variable long character string, for example, several tens of characters, is encoded with fixed information, for example, 12 bits.

For details of the dictionary type algorithm, see, for example, (1) TCBell etc, "Text Compression", Prentice-Hall,
1990. section 8, Dictionary Techniques. On the other hand, the probability statistic type reflects the appearance probability of the next character from the appearance probability (including the conditional probability dependent on the immediately preceding character string) of the individual character (constant because it is one character) that appeared in the past. Encoding is performed with statistical (entropy) information (code length is variable). Therefore, it is based on FV (Fixed-Variable coding) coding in which the appearance probability of each character (fixed) is encoded with statistical information (variable).

The details of the stochastic algorithm are described in (2) TCBell etc, "Text Compression", Prentice-Hall,
1990.section 5, From Probabilities To Bits. Among these, Huffman coding (Huffman coding) and arithmetic coding (Arithmetic codin
g) There is.

Further, the details of the context model that takes in the subordination relation of the immediately preceding character string are described in (3) TCBell etc, "Text Compression", Prentice-Hall,
1990.Section 6, Context modeling. In this case, the subordination of the previous string is
This is different from the dictionary type, which takes an infinitely long string with at most a few characters.

[0008] Therefore, in the dictionary type coding and the probability statistical type coding, the method of catching data that appeared in the past is the character string itself in the dictionary type coding, and the probability statistical type coding is the occurrence probability. The encoding part is fixed length in dictionary type encoding,
The compression mechanism is radically different in that stochastic coding is of variable length (in principle).

Here, multi-value arithmetic coding will be described as a universal coding method for a data stream of a byte stream such as English text. Broadly speaking, two types of arithmetic coding have been proposed. 2
Value arithmetic coding and multi-value arithmetic coding. The difference between the two is
The unit of data to be encoded by binary arithmetic encoding is 0
Or, it is a binary corresponding to one bit, and the unit of data to be encoded by the multi-level arithmetic coding is a difference between byte-compatible multi-level coding in which, for example, 8 bits are grouped together. A typical example of binary arithmetic coding is a QM coder used in JBIG entropy coding, which is a standardized method for binary image compression in CCITT / ISO. For example, see (4) Yasuda, "International Standards for Multiprocessor Media Coding: Chapter 3, Arithmetic Coding", Maruzen, pp.68-82,1991,6.

A typical example of multi-level arithmetic coding is WI
There is TTEN, ABRAHANSON encoding. For example, (5) IHWitten et al., "Arithmetic Coding for Data Com
pression '', Communications of Association for Comput
ing Machinery, 30 (6), 520-540, 1987.7 (6) DMAbrahamson, `` An Adaptive Dependency source
Mode for DataCompletion) Communications of Ass
ociation for Computing Machinery, 32 (1), 77 -83,198
Details are given in 9.1. Of these, only binary arithmetic coding has been put to practical use because it is suitable for images. However, since multi-level arithmetic coding provides the highest compression performance, its practical use is expected.

As shown in FIG. 67, the processing of the probability statistical coding is performed by using the appearance frequency model unit 400 and the entropy coding unit 4.
02. The appearance frequency model unit 400 takes in the input character and the character string (context) immediately before the input character, and obtains the appearance frequency of the input character in consideration of the dependency. The entropy coding unit 402 dynamically creates a code based on the number of appearances obtained by the appearance frequency model unit 400 and performs variable length coding.

More specifically, when a character string abc composed of three characters a, b, and c as shown in FIG. 68A is taken as an example, the relationship with the immediately preceding character string is shown in FIG. It is represented by such a tree structure. The appearance frequency model unit 400 counts the number of appearances each time a character string that passes through a character at each node of the tree structure appears, thereby determining the dependency on the immediately preceding character string,
For example, a conditional probability is determined. There are two methods of collecting contexts for obtaining the dependency between the input character and the immediately preceding character string: a method of collecting fixed-order contexts and a method of collecting blend contexts. Here, the number of characters in the context is called an order.

The method of collecting a fixed-order context is a method of fixing the number of characters in the context. For example, in the secondary context, the appearance frequency model unit 400 determines that the immediately preceding two characters x1, x2
Is collected, and the dependency of the character y appearing after the immediately preceding characters X2 and X1, for example, the conditional probability ρ (y
| X1, x2) is passed to the entropy coding unit 402. Here, y is the target input character, and x1 and x2 are the first character and the second character immediately before, respectively.

The method of collecting the blend context is a method in which the order of the context is mixed. In the case of a fixed-degree context, if the previous character string is difficult to appear, the estimation of its dependency becomes uncertain. Conversely, when the immediately preceding character string is likely to appear, the accuracy of the estimation of the dependency increases, leaving the possibility that the degree can be further increased. In general, the higher the context in which the immediately preceding character string becomes longer, the easier it is to catch the bias between characters, and a higher compression ratio is obtained.

However, data having a weak correlation in a higher-order context has a worse compression ratio. An attempt to solve such a problem is a blend context in which orders are mixed.
In the method of blending context, the order is not fixed, and if it is easy to appear depending on the individual context, the dependency is derived at a high order, and if it is difficult to appear, the dependency is derived at a low order. Things.

Based on the number of appearances determined by the appearance frequency model unit 400, typical encodings of the entropy encoding unit 402 for giving a dynamic code include arithmetic encoding,
There are dynamic Huffman coding, self-organizing coding, and the like. In arithmetic coding, a code is calculated from the appearance probability of each character, so that a code can be assigned even at 1 bit / character or less.
It is said that the coding efficiency is the highest.

FIGS. 69 (A) to 69 (C) show a multi-value arithmetic coding procedure in which a character string of an input character in units of a plurality of bits, for example, bytes, is associated with one point of a number line [0, 1]. 1
It is represented by two signs. Here, for simplicity, a,
Consider a character string consisting of the four characters b, c, and d. First, as shown in FIG. 69A, the appearance frequency of each character is obtained. Here, the appearance frequency is a probability obtained by dividing the number of appearances of each character by the total number of appearances. For example, the appearance frequency of the character a is 0.15, the appearance frequency of the character b is 0.25, and the appearance frequency of the character c is 0.3.
5. Assume that the appearance frequency of the character d is 0.25.

Next, regarding the appearance frequency in FIG. 69 (A), the characters are rearranged in descending order of frequency, and the cumulative appearance frequency is calculated as shown in FIG. 69 (B). The cumulative frequency is the sum of the frequencies of appearance of characters having a lower rank than the character of interest.
That is, the cumulative appearance frequency of the first character c having the highest appearance frequency is the sum of the appearance frequencies of the characters b, d, and a.
It becomes 5. Similarly, the cumulative appearance frequencies of the remaining characters b, d, and a are 0.40, 0.15, and 0.0.

If a character c is input in this state, for example, as shown in FIG. 67 (C), the cumulative appearance frequency freq [c of the input character c in the coding section [0, 1] defined by the number line ] = 0.35 and the cumulative appearance frequency cum fr
A new section 404 is obtained from eq [c] = 0.65. Specifically, the upper end (high level) H1 of the coding section [0, 1] defined by the number line, and the lower end (low level) L1
= 0, the section width W1 = 1, the cumulative appearance frequency freq [c] of the input character c = 0.35 and the cumulative appearance frequency cum
From freq [c] = 0.65, the upper end (high level) H2, lower end (low level) L2, and section width W2 of the new section 404
Ask for. That is, the lower end L2 of the new section 404 is calculated from the lower end (low level) L1 of the previous section and the section width W1 as follows: L2 = L1 + W1 × cum freq [c] = 0.0 + 1.0 × 0.65 = 0.65 Become. The width W2 of the new section 404 is W2 = W1.times.freq [c] = 1.0.times.0.35 = 0.35. The upper end (high level) H2 of the new section 404 is H2 = L2 + W2 = 0. 65 + 0.35 = 1.00. At this time, since the character c is input, the number of occurrences of the character c is incremented by one, and the total number of occurrences is incremented by one, and the appearance frequency and the cumulative appearance frequency of the characters a, b, c, and d are accordingly increased. Is updated. However, for the sake of simplicity, the description will be made with the appearance frequency and the cumulative appearance frequency shown in FIGS. 69 (A) and (B) fixed.

Next, when the character a is input, the previous section 4
04 is regarded as a new section [0, 1], and this section 404
, The appearance frequency of the input character a, freq [a] = 0.
15 and the cumulative appearance frequency cum freq [a] = 0.15
, The lower end (low level) L3 of the new section 406,
The section width W3 and the upper end H3 are obtained as follows. L3 = L2 + W2.times.cum freq [a] = 0.65 + 0.35.times.0.15 = 0.7025 W3 = W2.times.freq [a] = 0.35.times.0.15 = 0.0525 H3 = L3 + W3 = 0 .7025 + 0.0525 = 0.7550 Next, when the character d is input, the previous section 406 is regarded as a new section [0, 1], and the appearance frequency freq [d ] = 0.25 and the cumulative appearance frequency cum freq [d] = 0.4, the lower end (low level) L4, the section width W4, and the upper end H4 of the new section 408 are obtained as follows.

L4 = L3 + W3 × cum freq [d] = 0.725 + 0.0525 × 0.40 = 0.235 W4 = W3 × freq [d] = 0.0525 × 0.40 = 0.0210 H4 = L4 + W4 = 0.7235 + 0.0210 = 0.7445 If this input character d is the last character, section 408
, For example, an arbitrary value between the upper end and the lower end of the section 408 is output as an arithmetic code. In particular,
The previous section 406 was normalized to [1, 0] to obtain a threshold 1 /
The divided section is divided into four divided sections by 4, 1/2, and 3/4, and the upper end H4 and the lower end L4 of the final section 408 are coded to which of the normalized previous sections 406 belongs. For example, the previous section 406 is encoded under the following conditions.

Bit 1 is set when the value is 1/2 or more, and bit 0 is set when the value is less than 1/2. Bit 1 is set when the value is in the range of 1/4 or more and 3/4 or less, and bit 0 is set when the value is out of the range. In the case of the last section 408, since H4 = 0.7445, the bit becomes "1" under the condition, and the bit becomes "1" under the condition. Since L4 = 0.7235, the bit is "1" under the condition, and the bit is "1" under the condition. Therefore, the arithmetic code of the character string "cad" is "1111".

In the actual processing, when characters are input and encoded, the number of occurrences, the cumulative number of occurrences, and the total number of occurrences of the characters are obtained. Then, the number of appearances is divided by the total number of appearances to determine the cumulative appearance frequency, and the cumulative number of appearances is divided by the total number of appearances to determine the cumulative occurrence frequency. In this sense, the appearance frequency can be referred to as the number of appearances normalized by the total number of appearances, and the cumulative appearance frequency can be referred to as the cumulative number of appearances normalized by the total number of appearances.

According to such multi-level arithmetic coding, the character string having a higher appearance frequency has a larger width of the last section and can be represented by a shorter code, so that the data amount is compressed. Further, in this method, since there is no restriction on the minimum unit in the bit representation after coding and the minimum unit can be set to 1 bit or less, highly efficient compression can be performed. FIG. 70 is a block diagram of a conventional arithmetic encoding data compression apparatus. 70, the data compression device includes an appearance frequency order rearranging unit 41.
0, counter 412, frequency data storage 414, dictionary 4
16, an arithmetic coding unit 418. In this example, the number of appearances freq [] is used instead of the frequency of appearance, and the number of appearances cum f
req [] is used. The dictionary 416 may be included in the frequency data storage 414.

The counter 412 counts the number of appearances freq [] of the input character, and furthermore, the cumulative number of characters cum freq [] and the total number of appearances freq [].
[0] is obtained and stored in the frequency data storage unit 414. Every time a character is input, the frequency order rearranging unit 410 rearranges all the characters in the frequency data storage unit 414 in descending order of the number of appearances freq [], and sets the number of appearances freq [] in correspondence with the registration number indicating the order. Cumulative appearance number cum freq []
Is stored. At the same time, the symbols in the dictionary 416 are rearranged in the order of the number of appearances in the frequency data storage unit 414 in accordance with the registration number indicating the order.

When the arithmetic encoding unit 418 receives a registration number indicating the order from the frequency order rearranging unit 410 by referring to the dictionary 416 using the input character k, the arithmetic encoding unit 418 refers to the frequency data storage unit 414 based on the registration number, and Number of occurrences fr of k
eq [k], the cumulative appearance number cum freq [k], and the total appearance number cum freq [0] are obtained. Then, the number of appearances freq [k] and the cumulative number of appearances cum freq
[K] is divided by the total number of appearances cum freq [0] to calculate the appearance frequency and the cumulative appearance frequency, and a new section is obtained based on the calculated frequencies.

The operation of FIG. 70 will be described. The frequency order rearranging unit 410 searches the dictionary 416 using the input character k, obtains a registration number indicating the order of the number of occurrences, and calculates the arithmetic encoding unit 418.
Output to The arithmetic encoding unit 418 refers to the frequency data storage unit 414 based on the registration number (order) from the frequency order rearranging unit 410, and displays the number of appearances freq [k] of the input character k,
Cumulative appearance number cum freq [k] and total appearance number cu
Find m freq [0].

Then, the number of appearances freq [k] and the cumulative number of appearances cum freq [k] are calculated as the total number of appearances fr fr [k].
eq [0] to calculate the appearance frequency and the cumulative appearance frequency. Further, based on the previous section width Wk-1, the upper end (high level) Hk and the lower end (low level) Lk of the new section Wk are calculated. If it is the last character, an arbitrary value in the new section is output as a code.

On the other hand, the counter 412 indicates the number of appearances freq [k] of the input character k and the number of cumulative occurrences cum freq [k].
And the total number of occurrences cum freq [0] is incremented by one, and the value corresponding to the registration number in the frequency data storage unit 414 is updated. Subsequently, the frequency sorting unit 41
0 sorts the frequency data storage unit 414 and the dictionary 416 in the order of the updated appearance frequency freq [].

The flowchart of FIG. 71 is a process of multi-level arithmetic coding. In this example, a single-history of an appearing character without considering the history with the character immediately before it, that is, a case of estimating a subordinate relationship with a zero-order context is assumed. I have. Here, the initial values of the coding section are: upper end H0 = 1, lower end L0 = 1, section width W0
= 1.0. i is the order (registration number) of the dictionary sorted in descending order of the number of appearances, and 1, 2, 3,.
Take the value of A.

Further, freq [i] is the number of occurrences of the order (registration number) i of the dictionary sorted in descending order of the number of appearances, and cum freq [i] is the cumulative number of occurrences of the order i of the dictionary. . Furthermore, since I = 1, 2, 3,..., A, cum freq [1] is the cumulative appearance number of the first place, and cum freq [A] is the cumulative appearance number of the lowest A Represents Furthermore, the coding section is cum fr
It is normalized by eq [1] = 1, cum freq [A] = 0.

First, in step S1, the following initial values are set. All single characters are assigned to the array i of the dictionary D. D (i) = i where i = 1, 2, 3,..., A, where A is the number of alphabets, symbols, etc. and takes a maximum of 256.

An order (registration number) i is assigned to each character. I (i) = i where i = 1, 2,...,
A Initialize the number of appearances of all characters. freq (i) = 1, where i = 1, 2,.
A Initialize the cumulative number of all characters.

Cum freq (i) = A-1 After the initial values have been set, the first character k from the input character string of the source data is input in step S2. Next, the process proceeds to step S3, where the dictionary is searched by the input character k to determine the rank j as the registration number. This is j =
It is expressed as I (k). Further, the cumulative appearance number cum freq [j] is obtained from the rank j with reference to the list table of the cumulative appearance number corresponding to the code section. This is i = D (j)
Is expressed as

Then, the rank j searched from the dictionary is arithmetically encoded. Arithmetic encoding of rank j means that the cumulative appearance count cum freq [j] of rank j is calculated by the total appearance count cum freq
By dividing by [0], the cumulative appearance frequency is obtained, and a new section is obtained from the previous section width and the lower end (low level). Here, if the input character is the last character of a predetermined number of bytes serving as a coding unit, an arbitrary value in the new section is output as a code.

Next, in step S4, the input character k in the order j
If the characters having the same number of occurrences are arranged in the higher order, the character located at the top order, the number of occurrences, and the cumulative number of occurrences are exchanged. First, among characters having a rank smaller than the rank j, a rank r that is one level lower than the j-th character k and a character having a different number of appearances freq [k] is obtained. That is, freq [r]!
Find the maximum rank r such that = freq [j]. here! Means different.

Further, the rank of the list of the number of appearances of the section corresponding to the r-th character in the dictionary is I (r) = s, and the rank of the list of the cumulative number of appearances of the corresponding section is D (r). ) = T. Therefore, the j-th and r-th characters are exchanged in the dictionary, and the j-th and s-th occurrence numbers are exchanged in the section list table of the number of appearances.
Further, the j-th and t-th cumulative appearance numbers are exchanged in the section list table of the cumulative appearance number.

That is, the replacement of the number of appearances I (j) = s, the replacement of the cumulative number of occurrences D (j) = t, the replacement of dictionary order I (r) = j, and the replacement of dictionary characters D (r) = D
Perform (j). For example, if there are higher ranks j-1, j-2, and j-3 having the same number of appearances as the rank j of the input character k, the highest r = j-3 is obtained as the replacement destination, and the rank r And the character of rank j, the number of appearances, and the number of cumulative occurrences are exchanged.

Subsequently, in step S5, the number of appearances freq [k] of the replaced input character k is incremented by one, and further, the cumulative number of occurrences cum f of rank r + 1 or more is added.
Increment req [] by one. Of course, the total number of occurrences cum freq [0] is also incremented by one.
Then, returning to step S2 again, the next character is input,
Repeat the same process until there is no more source data.

The flowchart of FIG. 72 is a multi-level arithmetic coding for estimating a double history of appearance characters in consideration of the history with the immediately preceding character, that is, estimating the subordination relationship with the primary context (reference document). (5) See DMAbrahamson). In the processing of the double history, a dictionary is created for each character in consideration of a combination of two consecutive characters, and the appearance frequency of the character following each character is registered in the dictionary for each second character. I do. For example, in the dictionary of the character a, characters appearing after the character a, such as b for ab and c for ac, are registered, and the number of appearances of each character following a is obtained and registered.

Therefore, by setting the initial value in step S1, D (p, i) is assigned to the array of the dictionary D for each character i following the character p. Where p, i = 1,
2, ..., A, where A is the size (number) of the alphabet
It is. The number assigned to each character is I (p, i)
And Here, p, i = 1, 2,..., A. The other points are the same as the single history processing in FIG.

As entropy coding other than arithmetic coding in which dynamic coding is performed using the tree structure of FIG. 68 (B), dynamic Huffman coding, spray coding which is a kind of self-organizing coding, and the like. There is also. Dynamic Huffman coding is described in, for example, (7) DEKnuth, "Dynamic Huffman Coding", Journal of A
logorithms 6, 163-180 (1985) JSVitter, "Design an
d Analysis of Dynamic Huffman Codes ", Journal of AC
M, 34, 4, 825-845 (1987) (8) Tomohiko Uematsu, "Introduction to Document Data Compression Algorithm",
Chapter 5, CQ Publishing Company (1994) Patents: Reference No. 94-11634, provides further details. Spray coding is based on (9) DWJones, "Application of Splay Tree to Data Co.
mpression ", Commun. of ACM, 31, 8, 996-1007 (1987) Patent: Reference No. 94-01147.

[0043]

However, in such a conventional probability-statistic type encoding, as shown in FIG. 67, after the appearance frequency is obtained by the appearance frequency model unit 400,
In order to perform dynamic encoding in the entropy encoding unit 402,
There was a problem that processing took time. When the blend context is processed by the appearance frequency model unit 402,
Since dependency relations such as conditional probabilities were sequentially determined from a higher order to a lower order, processing time was greatly increased, and it was not practically used.

Accordingly, an object of the present invention is to provide a data compression device and a data decompression device capable of efficiently performing probability statistical coding and decoding of multivalued data in byte units to reduce the processing time to a practical level. With the goal. Also, in conventional encoding, mainly English texts, for example, AS
It is considered that the encoding is performed based on the byte unit in the byte stream direction for the CII 8-bit code.

For this reason, unicode or Japanese code having a 2-byte structure currently used as an international language code, full-color image data having a 3-byte RGB format,
For data having a word structure of a plurality of bytes such as a program code such as a byte or 8 bytes, the conventional encoding method is not very efficient. That is, in the conventional encoding method, data having a word structure of a plurality of bytes is processed in units of bytes, so that it takes a long time to process. For example, an example in which encoding in bytes is simply extended to words If one byte is extended to two bytes,
The amount of data, the processing time, and the memory capacity for obtaining a sufficient compression ratio are all 256 bytes for 1 byte, which is 256 times impractical.

Accordingly, the present invention provides a data compression apparatus and a data decompression apparatus which can efficiently perform encoding and decoding by a dictionary type and a probability statistical type in a byte stream direction in a word stream direction, thereby reducing a processing time to a practical level. It is intended to provide a device.

[0047]

FIGS. 1, 2, 3 and 4 are diagrams for explaining the principle of the present invention. According to the present invention, there is provided a data compression device and a data decompression device capable of efficiently performing probability statistical coding and decoding of multivalued data in byte units to reduce the processing time to a practical level.

( Pipeline Processing in Fixed-Order Context ) First, as shown in FIG. 1A, the present invention provides a data compression apparatus for encoding an input character according to the conditional ease of the immediately preceding character string. The data compression device includes an appearance frequency model unit 10
And the entropy encoder 12 are connected to the pipeline controller 1
5 causes a pipeline operation.

That is, the appearance frequency model unit 10 searches the frequency data storage unit in which the conditional easiness of the input character connected to the immediately preceding character string is registered for each character by the input character, and obtains the conditional easiness. The frequency data storage unit is registered and updated. The entropy encoding unit 12 encodes the input character according to the conditional ease determined by the appearance frequency model unit 10. The pipeline control unit 15 causes the appearance frequency model unit 10 to input a character to determine the conditional ease of appearance, and at the same time, causes the entropy encoding unit 12 to entropy encode the conditional ease of appearance determined for the immediately preceding input character. Perform a pipeline operation.

The pipeline of the appearance frequency model unit 10 and the entropy coding unit 12 can shorten the processing time and speed up the probability statistical coding. The present invention also provides
Provided is a data restoring apparatus for restoring a character based on an input code and conditional ease of a character string immediately before being restored. In the data restoration apparatus, the pipeline control unit 45 causes the entropy decoding unit 46 and the appearance frequency model unit 48 to perform a pipeline operation.

The entropy decoding unit 46 obtains conditional ease of appearance from the input code and the character string immediately before being restored. The appearance frequency model unit 48 retrieves the conditional easiness of the input character connected to the immediately preceding character string by the conditional easiness obtained from the input code in the frequency data storage unit registered for each character, and restores the character. At the same time, the frequency data storage unit is registered and updated.

The pipeline control unit 45 determines the conditional ease of appearance of the code input by the entropy decoding unit 46 and, at the same time, restores the appearance frequency model unit 48 from the conditional ease of appearance of the previous code. Perform a pipeline operation to By making the entropy decoding unit 46 and the appearance frequency model unit 48 into a pipeline, the processing time can be shortened and the probability statistical decoding can be speeded up.

( Parallelization of Part of Blend Context ) As shown in FIG. 1B, the present invention also provides a data compression apparatus and a data decompression apparatus in which the appearance frequency model part of each order is parallelized in a blend model. Provide equipment. First, a data compression device for blend contexts encodes a character from a high-order context consisting of a long character string immediately before to a low-order context consisting of a short character string with ease of conditional output. For example, appearance frequency model units 88, 90, and 92 are provided in parallel for each of the 0th, 1st, and 2nd orders. The appearance frequency model units 88, 90, and 92 for each order have a frequency data storage unit that registers, for each order, the conditional easiness of the input character connected to the immediately preceding character string. Is searched by referring to the frequency data storage unit, and the frequency data storage unit is registered and updated.

The appearance frequency model units 88, 90,
Subsequent to 92, an entropy encoding unit 12 is provided, and the conditional ease of output of each order output from the appearance frequency model units 88, 90, 92 for each order is converted from a higher order context to a lower order context. In order. Further, the pipeline processing is realized by providing a pipeline control unit 85 that performs a pipeline operation of the appearance frequency model units 88, 90, and 92 and the entropy encoding unit 12 for each order.

The data restoration apparatus for the blend context restores characters in order from a higher-order context consisting of an input code and a long character string just before being restored to a lower-order context consisting of a short character string. The parallelization of the data restoration apparatus is performed for each order following the entropy decoding unit 46, for example, the 0th order,
Primary and secondary appearance frequency model units 130, 132, 13
4 are provided in parallel.

The entropy decoding unit 46 obtains conditional ease of order from the higher-order context to the lower-order context based on the input code and the character string immediately before each degree already restored.
Appearance frequency model units 130, 132, 134 for each degree
Has a frequency data storage unit in which the degree of conditional appearance of the input character connected to the immediately preceding character string is registered for each degree. From the degree of conditional appearance of the own degree obtained by the entropy decoding unit 46, the frequency data is stored. Characters are restored in parallel by referring to the storage unit, and the frequency data storage unit is registered and updated.

For the pipeline construction of the data restoration apparatus, a pipeline control unit 125 for performing a pipeline operation of the entropy decoding unit 46 and the appearance frequency model units 130, 132, 134 for each order is further provided. ( Parallelization of entire blend context ) Further, as shown in FIG. 2 (C), the present invention provides a data compression device in which, in a blend model, each of an appearance frequency model unit and an entropy encoding unit is parallelized for each degree. And a data restoration device.

The data compression apparatus is provided with an appearance frequency model unit 16 for each order, for example, for each of the 0th, 1st and 2nd orders.
6, entropy coding section 168 and code merging section 1
70 are provided independently to parallelize. The appearance frequency model unit 166 for each order has a frequency data storage unit in which the degree of conditional appearance of the input character connected to the immediately preceding character string is registered for each order, and the input character connected to the immediately preceding character string in each order. Is searched in parallel with reference to the frequency data storage unit, and the frequency data storage unit is registered and updated.

The entropy encoder 168 for each order
Performs entropy encoding in parallel with the conditional ease of input characters connected to the immediately preceding character string in each degree output from the appearance frequency model unit 166. Code merging section 170
Collectively outputs the codes from the entropy coding unit 170 of each order. Pipelining of data compression equipment
A pipeline control unit 175 is provided for operating a plurality of appearance frequency model units 166 and a plurality of entropy encoding units 168 provided for each order in a pipeline for each order.

On the other hand, the data decompression device
1. Parallelization is performed for each order by an entropy decoding unit 214 provided independently for each order and a plurality of appearance frequency model units 216 provided independently for each order. The code separation unit 211 separates the input code into codes of each order. The entropy decoding unit 214 for each order performs a process of obtaining conditional ease of output for each order based on the code of each order separated by the separation unit 211 and the character string immediately before each order already restored. Perform in parallel for each order.

The appearance frequency model unit 216 for each order has a frequency data storage unit in which the conditional ease of input characters connected to the immediately preceding character string is registered for each order, and the entropy decoding unit for each order. A process of retrieving the frequency data storage unit and restoring the frequency data storage unit and updating the registration of the frequency data storage unit is performed in parallel for each order based on the conditional ease of each order obtained in 214.

The data restoration apparatus is pipelined as follows:
A pipeline control unit is provided for operating the entropy decoding unit 214 and the appearance frequency model unit 216 for each order in a pipeline operation for each order. 1 (A) (B) and 2
The appearance frequency model part of the data compression device and the data decompression device of (C) searches, registers, and updates the rank, the number of occurrences, the cumulative number of occurrences, and the total number of occurrences in context as information for guiding conditional ease of appearance. I do. The entropy coding unit and the entropy decoding unit perform arithmetic coding and arithmetic decoding.

Further, the pipeline control section determines the order, the number of appearances,
The reading, addition, and writing of the cumulative number of occurrences and the total number of occurrences in context are pipelined for each piece of information. Further, the pipeline control unit performs operations of the upper end and the lower end representing the arithmetic code space in parallel. (Encoding in the Word Stream Direction) According to the present invention, encoding and decoding by the dictionary type and the probability statistical type in byte units in the word stream direction are efficiently performed,
A data compression device and a data decompression device capable of reducing the processing time to a practical level are provided.

FIG. 3 is a diagram for explaining the principle of encoding and decoding in the word stream direction. FIG. 3A is a data compression apparatus that encodes word data composed of a plurality of bytes in units of words, and a byte array that separates the word data into units of bytes and converts them into byte data strings at the same byte positions. A plurality of byte encoding units 502 for encoding the input byte data at each byte position obtained by the conversion unit 500 and the byte array conversion unit 500 with conditional ease with the byte data which has been encoded; A merging unit 503 outputs the codes encoded by the plurality of byte unit encoding units 502 at a time.

Two-byte Unicode, Japanese code, R
In encoding for compressing a program code such as full-color image data of 3 bytes of GB, 4 bytes or 8 bytes, there is a strong correlation between byte data at the same byte position in word data. The present invention performs byte-by-byte encoding in the word stream direction, as opposed to byte-by-byte encoding in the conventional byte stream direction.Therefore, the correlation at the same byte position in the word can be incorporated into the encoding. High compression efficiency can be expected.

As the byte unit encoding unit 502 , a dictionary encoding unit or a probability statistical encoding unit is used. When the byte unit encoding unit 502 is a probability statistical encoding unit, it is configured by an appearance frequency model unit, an entropy encoding unit, and a pipeline control unit. Here, in the case of the context fixed in which the number of bytes of the context is fixed, the appearance frequency model unit has a frequency data storage unit in which the conditional ease of input byte data connected to the immediately preceding byte data string is registered for each byte data. Then, the conditionality of the input byte data connected to the immediately preceding byte data string is searched for by referring to the frequency data storage unit, and the frequency data storage unit is registered and updated.

The entropy encoding unit encodes the input byte data according to the conditional ease determined by the appearance frequency model unit. The pipeline control unit causes the appearance frequency model unit and the entropy coding unit to perform a pipeline operation. When the byte unit encoding unit 502 is a probability statistical encoding unit and a blend context in which the number of characters immediately before is fixed and variable is used, the appearance frequency model unit includes A frequency data storage unit registered for each degree from the immediately preceding higher-order context (long byte data string) to the lower-order context (shorter immediately preceding byte data string); Is provided for each order in order to retrieve the conditional ease of input byte data connected to the immediately preceding byte data string by referring to the frequency data storage unit and to perform a process of registering and updating the frequency data storage unit.

The entropy encoding unit encodes conditional ease of each order of the plurality of appearance frequency model units in order from a higher-order context to a lower-order context. Then, the pipeline control unit causes the plurality of appearance frequency model units and the entropy coding unit to perform a pipeline operation. As described above, for the encoding in the word stream direction, a data compression device in which the processing time is reduced to a practical level can be provided by achieving parallelization and pipelining.

A data decompression device for decompressing word data from a code coded by the data compression device shown in FIG. 3A converts a code at each byte position of word data from an input code string as shown in FIG. 3B. A code separating unit 530 for separation, a plurality of byte-based decoding units for decoding byte data according to the conditional output of the input code of each byte position separated by the code separating unit 530 and the already restored byte data at the same byte position And a switching output unit 533 that outputs the word data by arranging the byte data restored by the plurality of byte unit decoding units 532 in the order of the byte positions.

This data decompression device also corresponds to the data compression device, and the byte unit decoding unit 532 is a dictionary type decoding unit or a probability statistical type decoding unit. The probability statistical decoding unit
In the case of a fixed context, an entropy decoding unit that determines conditional ease of input from the input code and the immediately preceding byte data string that has already been restored, the conditional ease of input byte data connected to the immediately preceding byte data string in each byte An appearance frequency model that has a frequency data storage unit registered for each data, retrieves the frequency data storage unit based on the conditional ease determined by the entropy decoding unit, restores byte data, and registers and updates the frequency data storage unit And a pipeline control unit that performs a pipeline operation of the entropy decoding unit and the appearance frequency model unit.

In the case of the blend context, the probability statistic type decoding unit uses the input code and the already restored higher-order context to extract the higher-order context based on the immediately preceding byte data string in each order of the lower-order context. And an entropy decoding unit having a frequency data storage unit in which the conditional ease of input byte data connected to the immediately preceding byte data string is registered for each degree in order of low order context from A plurality of appearance frequency model units provided for each order for performing a process of restoring byte data by referring to the frequency data storage unit from the conditional easiness of each order obtained in the above and registering and updating the frequency data storage unit; and It is composed of an entropy decoding unit and a pipeline control unit that performs a pipeline operation of a plurality of appearance frequency model units.

When arithmetic coding is employed as the probability statistical coding, the appearance frequency model section in the data compression device and the data decompression device includes the rank, the number of occurrences, and the cumulative occurrences as information for guiding conditional easiness. Search, register, and update the number and total number of occurrences in context. The entropy coding unit and the entropy decoding unit perform arithmetic coding and arithmetic decoding. (Switching between Encoding and Decoding in Word Stream Direction and Byte Stream Direction) FIG. 4 is a diagram illustrating the principle of adaptively switching encoding and decoding in byte units in the word stream direction and the byte stream direction.

FIG. 4A shows a data compression apparatus which encodes word data composed of a plurality of bytes in units of words, and includes a byte array conversion section 600, a word stream encoding section 602, a byte stream encoding section 604, It comprises a code switching unit 605. The byte array conversion unit 600 separates the input word data into byte data and then converts the input word data into a two-dimensional array for a plurality of words in the word stream direction and the byte stream direction, and converts this two-dimensional array for a plurality of words (an order of context). Hold.

The word stream encoding unit 602 converts each byte data of the input word in the word stream direction of the two-dimensional array by the byte array conversion unit 600 with the byte data at the same byte position of the encoded word data group. A word stream code is generated by encoding according to conditional ease. This corresponds to the data compression device in FIG.

The byte stream encoding section 604 encodes each byte data of the input word in the byte stream direction of the two-dimensional array by the byte array conversion section 600 according to the conditional ease with the preceding byte data to perform byte stream encoding. Generate This is equivalent to a conventional data compression device that takes in the byte data sequence as it is and encodes it in byte units.

The code switching unit 605 switches and outputs either a word stream code or a byte stream code. For example, the code switching unit 605 switches the code output based on the strength of the correlation between the word stream code and the byte stream code. Further, a shorter code length is selected from the word stream code and the byte stream code. The code switching unit 605 switches and outputs the word stream code and the byte stream code in byte units. The code switching unit 605 can also switch and output the word stream code and the byte stream code in word units.
Further, the code switching unit adds a switching code every time switching between the word stream code and the byte stream code.

The code switching unit 605 uses a switching code based on the frequency of occurrence of each past code as a switching code between the word stream code and the byte stream code. FIG. 4 (B)
FIG. 4 is a diagram illustrating the principle of a data decompression device that decompresses word data from a code by the data compression device of FIG.
A code separation unit 700, a word stream decoding unit 702,
Byte stream decoding unit 704 and output switching unit 706
It consists of.

The code separation section 700 receives the code string and
Separate word stream code or byte stream code at each byte position of the word. The word stream decoding unit 702 decodes the byte data according to the conditional ease of byte data at the same byte position that has already been restored and the word stream code separated by the separation unit 700. The byte stream decoding unit 704 decodes the byte data according to the conditional ease of the byte stream code separated by the code separation unit 700 and the previously restored preceding byte data. The output switching unit 706 combines the byte data decoded by the word stream decoding unit 702 or the byte stream decoding unit 704 in the order of the byte positions, and outputs the restored word data.

According to the encoding in the word stream direction in FIG. 3, 2-byte Unicode, Japanese code, RGB
With regard to encoding of 3-byte full-color image data, multi-byte word data such as 4-byte and 8-byte program codes, etc., an improvement in compression efficiency can be expected. However, depending on the word data, the encoding efficiency in the conventional byte stream direction may be higher in compression efficiency,
There are cases where it cannot be decided uniquely.

Therefore, the compression efficiency can be further increased by performing the word stream direction and the byte stream direction for the same byte data and adaptively selecting a code having a high compression efficiency.

[0081]

BEST MODE FOR CARRYING OUT THE INVENTION

<Table of Contents> 1. Pipeline processing in fixed order context 2. Parallelizing a part of the blend context 3. Overall blend context parallelization 4. Access to frequency data using hash structure 5. Coding and decoding in the word stream direction Adaptive switching process in word stream direction and byte stream direction Pipeline Processing in Fixed-Order Context FIG. 5 shows the basic configuration of a probability statistical data compression apparatus according to the present invention. As shown in FIG. 5A, the data compression device includes an appearance frequency model unit 10 and an entropy encoding unit 12, and the pipeline control unit 11 causes the appearance frequency model unit 10 and the entropy encoding unit 12 to perform pipeline processing. Controlled to do so.

The appearance frequency model section 10 inputs the first character of a character string prepared as source data, and searches for the conditional ease of input characters connected to the immediately preceding character string. That is,
The appearance frequency model unit 10 is provided with a frequency data storage unit that registers the conditional ease of input characters connected to the immediately preceding character string for each character, and searches the frequency data storage unit based on characters input from source data. Then, a conditional ease of appearance of the corresponding input character is obtained. If the conditional ease of input characters can be searched, the content of the frequency data storage unit is registered and updated. The entropy coding unit 12 performs entropy coding of the conditional ease of input characters obtained by the appearance frequency model unit 10. As the encoding by the entropy encoding unit 12, multi-level arithmetic encoding, dynamic Huffman encoding, or spray encoding, which is a type of self-organizing encoding, is performed.

FIG. 5B shows the appearance frequency model unit 10 and the entropy coding unit 1 by the pipeline control unit 11.
2 illustrates a pipeline operation. Here, the time axis is shown by time cycles T1, T2, and T3. Now, consider a case where characters a, b, a... Are input from source data. First, the leading character a is input to the appearance frequency model unit 10 in a T1 cycle, and the conditional appearance of the input character a is determined. In the next T2 cycle, the second letter b
At the same time but has been ease out conditionally input to the occurrence frequency modeling unit 10 is required, conditional Deeki of the first character a obtained in T1 cycle is given to the entropy coding unit 12, the frequency of the input character a Is performed.

In the same manner, the appearance frequency model unit 10 inputs one character to determine conditional ease of appearance, and at the same time, the entropy encoding unit 12 determines the entropy of the conditional ease of appearance previously determined by the appearance frequency model unit 10. The encoding is performed simultaneously and in parallel. As described above, since the appearance frequency model unit 10 and the entropy encoding unit 12 operate independently of the pipeline by the pipeline control unit 11, encoding of the character string by the probability statistical type can be performed at higher speed.

FIG. 6 shows an embodiment of the data device of the present invention in which arithmetic coding is applied to the data compression device of FIG. When arithmetic coding is applied, the data compression device includes a probability model unit 14, an arithmetic coding unit 16, and a pipeline control unit 15. The probability model unit 14 outputs, for example, a, a,
.. are input, and the probability (appearance frequency = normalized number of appearances) of the conditional appearance of the input character is calculated for each input character.

More specifically, the probability model unit 14 is provided with a frequency data storage unit. The frequency data storage unit is provided for each rank in accordance with the registration numbers or ranks arranged in the order of the number of appearances of each character. The character symbol, the number of appearances, and the cumulative number of appearances are registered, and the total number of appearances is stored. For this reason, the frequency data storage unit is searched for the input character a, a rank is obtained as a conditional probability corresponding to the input character a, and the rank is output to the arithmetic coding unit 16.

At the same time, the appearance frequency, the cumulative appearance frequency, and the total appearance frequency of the input characters used for calculating the appearance frequency and the accumulated appearance frequency in the arithmetic coding are output. The arithmetic coding unit 16 performs multi-level arithmetic coding. More specifically, arithmetic coding of a rank, which is a conditional probability of an input character output from the probability model unit 14, is performed. Arithmetic encoding of the rank of an input character, which is a conditional probability, means that the code space of the number line [1,0] is used as an initial space, and the appearance frequency and the cumulative appearance frequency corresponding to the rank of the input character in the previous code space are used. An assigned space for the input character is calculated, and an appropriate value indicating the calculated space is output as a code.

In the example of FIG. 6, a code is generated for each character and output as compressed data. Specifically, for the input character a, the first place is obtained by the probability model unit 14, and this is encoded by the arithmetic encoding unit 16 to obtain the code 0.
Outputs 1. For the next character b, the probability model unit 1
4, the second place is obtained. This is coded by the arithmetic coding unit 16 and the code 1001 is output. Further, as for the third character a, the first place is arithmetically coded and the code 01 is output, similarly to the first character.

In an actual apparatus, a code space corresponding to the order is determined in order from the first character, with a plurality of characters of a predetermined number of bytes, for example, a plurality of characters of 8 bytes as one group. A process of outputting the represented value as a code is performed. FIG. 6B shows the probability model unit 14 and the arithmetic coding unit 16 by the pipeline control unit 15.
This is the operation of the pipeline processing. First, in the T1 cycle, the leading character a is input, and the probability model unit 14 obtains the first place of the conditional probability. At this time, the arithmetic encoding unit 16 does not perform encoding because there is no previous ranking.

In the next T2 cycle, the second character b
Is input, the probability model unit 14 obtains the second place, and the arithmetic coding unit 16 performs the arithmetic coding of the first place of the input character a obtained in the previous cycle, and outputs the code 01. In the next T3 cycle, the same character a as the first character is input, and the first place is obtained by the probability model unit 14, and at the same time, the encoding of the second place of the character b obtained in the previous cycle T2 is performed. The arithmetic operation is performed by the arithmetic coding unit 16 to output a code 1001. In the same manner, a process of inputting one character in the probability model unit 14 to obtain a rank and simultaneously performing arithmetic coding of the rank obtained in the previous cycle by the arithmetic coding unit 16 are repeated.

FIG. 7 is a block diagram showing the detailed configuration of the data compression apparatus using the arithmetic coding shown in FIG. In FIG.
The data compression device includes a probability model unit 14, an arithmetic coding unit 16
, And both are pipeline-controlled by the pipeline control unit 15. The probability model unit 14 includes a rank read / update unit 18, a character read / update unit 20, and a frequency read / update unit 2.
2. Cumulative frequency read / update unit 24, total cumulative frequency read / update unit 2
6. Further, a frequency data storage unit 25 is provided. The arithmetic coding unit 16 has an upper end (high level) of the code space.
, A code space calculation unit 30 for calculating the lower end (low level) of the code space, and a code output unit 36.

First, frequency data as shown in FIG. 8, for example, is stored in the frequency data storage unit 25 of the probability model unit 14. In FIG. 8, the frequency data storage unit 25 registers the character symbols K <b> 1 to K <b> 256 in order of the largest number of occurrences 1 to 256. The order and character symbol registration structure correspond to the dictionary 416 shown in the conventional device of FIG. At the same time, characters K1 to K2 corresponding to ranks 1 to 256
The number of appearances f (1) to f (256) of 56 are registered.
Further, the cumulative appearance numbers cf (1) to cf (2
56) is registered. Thus, the frequency data storage unit 2
In No. 5, since the number of appearances and the cumulative number of appearances are registered, the appearance frequency is obtained as frequency = number of appearances / total number of appearances. The cumulative frequency is also obtained as cumulative frequency = cumulative number of appearances cf (s) / total number of appearances S0. In the ordinary arithmetic coding algorithm, the number of appearances is represented by freq (), and the cumulative number of occurrences is represented by cum freq (). In FIG. 8, f () and cf () are respectively used. It has been simplified.

Referring again to FIG. 7, the probability model unit 14
The order reading / updating unit 18 refers to the frequency data storage unit 25 every time the first character is input from the character string, and searches for a registered character corresponding to the input character. This character search result is provided to the character reading / updating unit 20. Character reading and updating unit 20
Is a frequency data storage unit 2 corresponding to the searched input characters.
5 is read as the dictionary registration number.

Further, the search result by the rank reading / updating unit 18 is also given to the frequency reading / updating unit 22 and the cumulative frequency reading / updating unit 24, and the number of appearances f () corresponding to the character searched from the frequency data storage unit 25 and the cumulative Read the number of appearances f (). Further, the total cumulative frequency reading / updating unit 26 reads out the total number of appearances cf (0) = S0 at that time. In this manner, the frequency data storage unit 2 corresponding to the input character is
The rank, the number of appearances, the cumulative number of appearances, and the total number of appearances of the input characters read from 5 are given to the arithmetic encoding unit 16.

When the frequency data based on the input characters has been read out and output, the order reading and updating unit 18 and the character reading and updating unit 2
The frequency data storage unit 25 is updated by each of the frequency read / update unit 22, the cumulative frequency read / update unit 24, and the total cumulative frequency read / update unit 26. Specifically, after adding 1 to the number of appearances of the input characters, the number of characters and the number of cumulative occurrences are rearranged in the descending order of the number of appearances.

The arithmetic coding unit 16 calculates the number of occurrences f (s), the cumulative number of occurrences cf (s), and the total number of occurrences cf (0) = S0 of the rank s of the input character output from the probability model unit 14 by using
The allocation space of the current input character with respect to the previous code space is calculated as a new code space. For example, for the input character having the rank s in FIG. 6, the number of appearances f (s) and the cumulative frequency f
(S) Further, assuming that the total number of appearances S0 is given from the probability model unit 14, the cumulative appearance number cf (s-1) of the rank s-1 which is one rank higher than the rank s of the input character Ks is read, and the input character Ks is read. The section width Wnew, the upper limit (high level) Hnew, and the lower limit (low level) Ln of the new section to be the assigned section of s
ew is calculated.

First, the previous section width W is calculated according to the following equation (1).
Old is calculated. In the case of the first character, the previous section width Wold uses the initial value 1.0. Wold = (Hold−Lold) +1 (1) where Wold is the previous section width (range) Hold is the previous upper end (high level) Lold is the previous lower end (low level) The section calculated by this equation (1) Based on the width, the upper end (high level) Hnew and the lower end (low level) Ln of the new section
ew is calculated by the following equations (2) and (3). Hnew = Lold + {Wold × cf (s-1) / S0} -1 (2) where Hnew is the current upper end (high level) cf (s-1) is the cumulative appearance of the next higher rank s-1 The number So is the total number of occurrences Lnew = Lold + {Wold * cf (s) / S0} (3) where Hnew is the current upper end (high level) cf (s) is the cumulative number of occurrences of the next higher rank s So Is the total number of appearances. The calculation of the high level (Hnew) in the code space according to the equations (1) and (2) is performed by the code space calculation unit 28. The calculation of the low level (Lnew) of the code space according to the expressions (1) and (3) is performed by the code space calculation unit 30.
Done in That is, the high level and the low level of the new code space are calculated in parallel by the two code space calculation units 28 and 30.

In order to simplify the division in the code space calculators 28 and 30, the reciprocal 1 / S of the total number of appearances S0 output from the total cumulative frequency reading / updating unit 26 in the reciprocal calculator 32 is calculated.
0 is calculated. By calculating the reciprocal 1 / S0, the speed can be increased by performing the division of the expressions (2) and (3) in the code space calculation units 28 and 30 in the form of multiplication. The high level and low level of the assigned space of the input character a calculated by the code space calculation units 28 and 30 are given to the arithmetic code derivation unit 34. The arithmetic code deriving unit 34 generates bit information indicating the upper and lower ends of the code space as shown in FIG. That is, as shown in FIG. 10, the previous encoding section is normalized as [1.0], and this section [1,0] is thresholded by 1/4, 1/2, 3/4.
Is set and divided into four sections.

For each of the divided sections [1, 0], an encoding is performed to determine which of the divided sections has the high level and the low level of the newly calculated input code allocation section. That is, the calculated high level and low level are as follows.
Encode with two conditions. <Condition> Bit 1 is set when the value is 1/2 or more, and bit 0 is set when the value falls below 1/2.

<Condition> Bit 1 is set if it is in the range of not less than 1/4 and not more than 3/4,
If it is out of this range, bit 0 is set. FIG. 10 shows the allocation range of the calculated coding section with respect to the normalized section [1, 0] by a thick line. First, the code space 38-1,3
8-2, 38 -3,38-4 the high level is 1.0, is when the low level is present in order from top to divided space. The output code in this case is 4 bits b0, b1, b
2 and b3 and take the values shown. For example, taking the code space 38-1 as an example, first, the high level is 1.0, and since it is 1/2 or more from the condition, b0 = 1.

Since the condition is outside the range of not less than 1/4 and not more than 3/4, b1 = 0. Next, for the low level, b2 = 1 because it is 1/2 or more from the condition, and b3 = 0 because it is outside the range of 1/4 or more and 3/4 or less from the condition. Code space 3
For 8-2, similarly, b0 b1 b2 b3 = 10
It becomes 11. For the code space 38-3, b0 b1 b
2b3 = 1001. Further, for the code space 38-4, b0 b1 b2 b3 = 1000. Although the high level is set to 1.0 for the code spaces 38-1 to 38-4, the same output code is obtained if the high level is 1 to 3/4.

In the code spaces 40-1 to 40-3, the high level is ／ and the low level is present in the lower divided space in order. In this case, the high-level output code b
All 0b1 are 11, and the low level code b2 b3 changes to 11,01,00. Code space 40-1,
40-2 is a case where the high level is １ ／ and the low level is present in the lower divided space in order. In this case, the high-level output code b0 b1 is 01, and the low-level output code b2 b3 is 01,00. Code space 4 already
4 is a case where both the high level and the low level exist in the range of 1/4 to 0, and in this case, the output codes b0 to b
All 3 are 0.

The lower and upper ends of the output code of FIG. 9 correspond to the range of the arithmetic code by comparison with the divided code space of FIG.
This is 4-bit information obtained by combining the four range determination bits. Next, the pipeline operation of arithmetic coding in the data compression device of FIG. 7 will be described with reference to the time chart of FIG. Here, FIG. 11 illustrates the pipeline operation from T1 to T1.
It is divided into nine time cycles. T1 to T9
The cycle is a normal case where the order of the input character is not the first place, and the T10 to T13 cycle is that the order of the input character is the first place.
In the case of the rank, the cycle of T14 to T19 is the case where there is no character of the higher rank having the same frequency as the rank of the input character.

First, the pipeline operation in the normal case of the T1 to T9 cycles will be described. First, the order reading / updating unit 18 of the probability model unit 14 inputs a character K to be encoded in the T1 cycle, and performs a process RK (meaning Read K) for reading out the order s by referring to the frequency data storage unit 25. The rank s of the input character K read in the T1 cycle is:
The frequency read / update unit 22, the cumulative frequency read / update unit 24, and the total cumulative frequency read / update unit 26 are further provided with the number of appearances f of the rank s in the T2 cycle as shown in the processes Rs, Rs, and R0.
(S), the cumulative appearance number cf (s) and the total appearance number S0 are read. In parallel with this, reading of the character Ks-1, the number of appearances f (s-1), and the number of cumulative occurrences cf (s-1) corresponding to the rank s-1 one rank higher than the rank s is performed by the processing Rs- As shown in FIG. 1, they are performed in parallel in the T3 cycle.

The result of reading the appearance data in the probability model unit 14 in the T1 to T3 cycles is provided to the arithmetic coding unit 16, and the high level and low level calculation processes of the code space are performed in parallel in the T3 to T5 cycles. Done. First, in the T3 cycle, for example, the code space calculation unit 28 performs addition and subtraction for calculating the previous section width Wold according to the above equation (1), and this result is also reflected on the code space calculation unit 30 that performs low level calculation. Let it.

At the same time, in the T3 cycle, the reciprocal 1 / S0 of the total number of appearances S0 is calculated by the reciprocal calculator 32 and supplied to the code space calculators 28 and 30. T4-T6
In the cycle, each of the code space calculation units 28 and 30 performs the calculation of the high level according to the equation (2) and the calculation of the low level according to the equation (3) in parallel. This T3-T
The calculation results of the high level and the low level of the code space over six cycles are provided to the arithmetic code deriving unit 34, and T7 to T1
The code [0101] is derived by comparison processing with the threshold value of the divided space shown in FIG. 8 in the 0 cycle. This derived code is the same as the code space 42-1 in FIG. The derived arithmetic code is output from the code output unit 36 as, for example, serial data.

On the other hand, in the probability model unit 14, after outputting the necessary frequency data to the arithmetic coding unit 16,
The update processing of the frequency data storage unit 25 is performed in cycles 3 to T9. First, in the T3 to T5 cycles, the highest rank having the same size as the number of appearances f (s) of the rank s of the input character K is obtained. That is, processing Rs-1, T3 to T5 cycle processing,
As shown in Rs-2 and Rs-3, the number of appearances is sequentially obtained from the top ranks, and the highest rank having the same appearance frequency as the rank s is searched.

This situation can be understood by referring to the frequency data storage unit 25 shown in FIG. Assuming that the order of the input characters is s, the number of appearances is f (s). Therefore, the higher ranks s1, s2, s3,.
And the number of occurrences f (s-1),
f (s-2) and f (s-3) are read out and compared with the number of appearances f (s) of the rank s.

Here, assuming that the number of appearances f (s-3) of the rank s-3 is larger than the number of appearances f (s) of the rank s, the rank having the same number of appearances f (s) is s2. .
Thereafter, at the time of updating for replacement, the number of appearances f (s) of the rank s is incremented by one in cycle T3. When the highest rank s-2 having the same value as the number of appearances f (s) of the rank s is obtained in this manner, the number of appearances and the cumulative number of characters corresponding to the rank s and the rank s-2 are obtained. Are performed in cycles T6 to T7.

Further, with respect to the cumulative appearance number, the cumulative appearance numbers cf (s-3) to cf (1) of the rank s-3 or higher are sequentially read and written to increment by one. This processing is represented by Ws-3... W1. Further, in the cycle T6, the total number of appearances S0 is incremented by one as shown in the processing W0. Thus, the update processing of the appearance data storage unit 25 accompanying the arithmetic encoding of the character K is completed.

The T10 to T13 cycles correspond to the case where the order of the input character is first. In this case, the frequency data storage unit 25 does not need to be replaced, and the rest is the same as the normal case of the T1 to T9 cycles. Further, T14 to T19
The cycle is a case where there is no higher rank having the same number of appearances as the number of appearances of the input character. In this case, the number of appearances f of the rank s is not changed for the rank s without replacing the characters.
The increment of (s), the respective increments of the cumulative appearance numbers cf (s-1) to cf (1) above the rank s-1 or higher, and the total appearance number S0 are also incremented. The processing of the arithmetic encoding unit 16 following the cycles T10 to T13 and T14 to T19 is the same as the encoding processing following the cycles T1 to T9.

FIG. 12 shows a basic configuration of a data decompression device for decompressing the original source data from the compressed data obtained by the data compression device of FIG. This data restoration device is shown in FIG.
As shown in (A), it is composed of an entropy decoding unit 46, an appearance frequency model unit 48, and a pipeline control unit 47. The entropy decoding unit 46 inputs codes one by one from the code string as compressed data, obtains conditional ease of appearance from the input code and the character immediately before being restored, Output. The appearance frequency model unit 48 has a frequency data storage unit that registers, for each character, the conditional easiness of the input character connected to the immediately preceding character string, and the conditional easiness of the input code output from the entropy decoding unit 46. To search the frequency data storage unit to restore characters.

The pipeline control unit 47 causes the entropy decoding unit 46 and the appearance frequency model unit 48 to perform a pipeline operation. Here, as the entropy decoding unit 46, arithmetic decoding, dynamic Huffman coding, and spray decoding, which is a kind of self-organizing coding, employed in the entropy coding unit 12 in FIG. , Dynamic Huffman decoding, splay decoding, and the like.

FIG. 12B shows an entropy decoding unit 46 and an appearance frequency model unit 4 by the pipeline control unit 47.
8 is a time chart of the pipeline operation of FIG. First T
A code is input in one cycle to perform entropy decoding for obtaining conditional ease. At this time, the appearance frequency model unit 48
Does not perform character restoration processing since the previous decoding result has not been obtained.

In the next T2 cycle, entropy decoding is performed by the entropy decoding unit 46 for inputting the next code and similarly obtaining conditional easiness. At the same time, the character a is restored by the appearance frequency model unit 48 by searching the frequency data storage unit based on the conditional ease determined in the previous cycle T1.
Similarly, the code is input and decoded by the entropy decoding unit 46, and at the same time, the character is restored and output from the decoding result obtained in the previous cycle by the appearance frequency model unit 48.

FIG. 13 shows a data decompression device for performing an arithmetic decoding for restoring characters from the compressed data obtained by the data compression device to which the arithmetic coding shown in FIG. 6 is applied. This data restoration device includes an arithmetic decoding unit 50, a probability model unit 52, and a pipeline control unit 51 as shown in FIG. Arithmetic decoding section 50 inputs codes 01, 1001, and 01 obtained by the data compression apparatus of FIG. 4 as compressed data in code units, obtains the order of the allocated code space obtained from the codes, and sends the order to probability model section 52. Output.

In this case, the first place, the second place, and the first place are sequentially output. The probability model unit 52 includes an appearance data storage unit having the same structure as that shown in FIG. 6 and arranged in the order of the number of appearances of each character in descending order, according to the order output from the arithmetic decoding unit 50. The frequency data storage unit is searched, and the corresponding characters are restored as a, b, and a. FIG. 13B is a time chart of the pipeline operation of the arithmetic coding unit 50 and the probability model unit 52 by the pipeline control unit 51. First, in the T1 cycle, the arithmetic coding unit 5
The code 01 is input to 0, and the first place is decoded. At this time, since the previous decoding order has not been input to the probability model unit 52, the character is not restored.

In the next T2 cycle, the second code 1001 is input to the arithmetic coding unit 50 to decode the rank 2, and at the same time, the frequency is calculated by the rank 1 obtained in the previous cycle T1 in the probability model unit 52. Search the data storage to find the letter a
To restore. Similarly, the arithmetic decoding unit 50 inputs a new code to decode the rank, and at the same time, the probability model unit 52 restores the character according to the previously decoded rank.

FIG. 14 shows details of the data restoration apparatus of FIG. This data restoration device includes an arithmetic decoding unit 50, a probability model unit 52, and a pipeline control unit 51.
The arithmetic decoding unit 50 includes a code input unit 54 and a cumulative value deriving unit 5
6. Code space calculation unit 5 for calculating the high level of the code space
8. Code space calculator 6 for calculating the low level of the code space
0, a reciprocal calculator 6 for calculating the reciprocal 1 / S0 of the total number of occurrences S0 used for calculating the high-level and low-level code spaces.
2. Cumulative frequency deriving unit 64 and reciprocal calculating unit 6 for calculating reciprocal 1 / Wold of range Wold used for cumulative frequency deriving
6 are provided.

The probability model unit 52 includes a rank deriving unit 6
8, an appearance data storage unit 70, a character reading / updating unit 72, a frequency reading / updating unit 74, a cumulative frequency reading / updating unit 76, a total cumulative frequency reading / updating unit 78, and a character output unit 80 are provided. Each time a code is input, the arithmetic decoding unit 50 calculates a cumulative occurrence number cf (code) given by the input code. The cumulative occurrence number cf (code) of this code is given by the following equation. cf (code) = [｛(Value−Lold) × S0｝ −1] / Wold (4) where “Value” is the cumulative value obtained from the code “Lold” is the lower end (low level) of the previous section. “Wold” is the section width of the previous section. (Range) This equation (4) defines the high level of the new space at the time of encoding,
Cumulative number c of rank s-1 which is one rank higher than the rank of the input character
This is the same as the expression (2) calculated based on f (s-1) and modified to the expression for calculating the cumulative appearance number cf (s-1). That is, the cumulative number c of the rank s-1 in the equation (2)
f (s−1) is replaced with the cumulative number of occurrences of the code cf (code), and the high level Hnew of the new space is calculated as the cumulative value Valu.
By substituting for e and deforming, equation (4) is obtained.

The cumulative value Value in the equation (4) is obtained by comparing with the threshold values 1/4, 1/2, and 3/4 set in the code section [1, 0] normalized from the input code. It is determined according to. <Condition> The low level of the new space obtained from the code is 1/2.
If this is the case, the accumulated value Value is set to 1/2.

<Condition> If the low level of the new space obtained from the code is not less than 1/4 and the high level is less than 3/4, the accumulated value Value is set to 1/4. Derivation of the cumulative value Value according to such a condition is performed by the cumulative value deriving unit 56. The calculation of the high level and the low level of the code space of the previously decoded code used for the calculation of the equation (4) is performed by the code space calculation units 58 and 60.
It is done in. In this case, the calculated values of the high level and the low level of the code space are basically the same as the above-described equations (2) and (3).

The cumulative frequency deriving section 64 calculates the cumulative occurrence number cf (code) of the code according to the above equation (4), and calculates the reciprocal of the lens Wold by the reciprocal calculating section 66 to the calculated cumulative occurrence number cf (code) of the code. The cumulative frequency of the code is derived by multiplying 1 / Wold. On the other hand, the probability model unit 52 calculates the cumulative number of appearances restored from the code output from the arithmetic decoding unit 50 in the rank deriving unit 68,
The cumulative number of occurrences of each character stored in the frequency data storage unit 70 is compared in order from the first place. The rank is derived by comparing the cumulative number of occurrences of this code with the cumulative number of occurrences from the higher rank stored in the frequency data storage unit 70.

When the rank is derived from the cumulative appearance frequency of the code by the rank deriving section 68, the rank is given to the character reading / updating section 72, and the character corresponding to the rank is stored in the frequency data storage section 70.
The character is restored by reading from the character string, and the character output unit 80 outputs the character. The frequency read / update unit 74 increments the number of appearances of the restored character by one. The cumulative frequency reading / updating unit 76 performs an update for incrementing the cumulative number of occurrences by one for higher ranks having a cumulative number of occurrences larger than the restored character. Further, when the rank of the same cumulative appearance number exists in the upper rank of the character of the restoration rank, the highest rank of the same cumulative appearance count and the restoration rank are replaced with respect to the character appearance count and the cumulative appearance count. Further, the total cumulative frequency reading / updating unit 78 increments the total number of appearances by one.

FIGS. 15 and 16 are time charts of the pipeline operation of the data restoration apparatus of FIG. Here, the T1 to T13 cycles in FIG. 15 are a normal case in which the order of the character restored from the code is other than the first place.
From the T14 cycle of No. 5 to the T24 cycle of FIG. 16, the order restored from the code is the first place. Further, cycles T25 to T38 in FIG. 16 are normal cases in which the order restored from the code is other than the first place, but there is no higher order of the same frequency in the restored order.

First, the normal restoration processing in the T1 to T13 cycles in FIG. 15 will be described. In the T1 to T2 cycles, a sign bit is input by the sign input unit 54, and the accumulated value derivation unit 56 derives the accumulated value by the above-described condition comparison processing. At this time, in the cycle before the T1 cycle, the calculation of the high level and the low level of the code space previously restored by the code space deriving units 58 and 60 has already been completed.

In the next T3 to T6 cycles, the cumulative frequency deriving unit 64 calculates the cumulative code appearance frequency according to the above equation (4). At this time, in the T4 cycle, the reciprocal calculator 66 calculates 1 / Wold in the equation (4).
Are performed in parallel, and the result is used for multiplication of T6. The cumulative occurrence number cf (code) of the code calculated in the T6 cycle is given to the probability model unit 52,
In the 7th to T13 cycles, the character is restored and the frequency data storage unit 70 is updated. In parallel with the restoration of the character and the update of the frequency data in the probability model unit 52, the arithmetic decoding unit 50 sets a high code space for the character being restored by the code space calculation units 58 and 60 in the T10 to T13 cycles. The calculation of the level and the low level are performed in parallel.

Before calculating the high level and the low level of the code space in the T10 to T13 cycles, T8
In the cycle, the reciprocal calculation unit 62 previously calculates the reciprocal 1 / S0 of the total number of appearances S0, thereby enabling the multiplication in the T12 cycle. Here, a description will be given of a process of restoring characters and updating frequency data in the probability model unit 52 in the T7 to T12 cycles. First, in the T7 to T10 cycles, the cumulative appearance number is read out in order from the one with the highest frequency registered in the frequency data storage unit 70, and the arithmetic decoding unit 50
The order in which the cumulative occurrence number cf (code) of the code sent from is located is derived from the magnitude relation between the read order and the cumulative occurrence number of the next higher rank.

That is, in the T7 to T10 cycles,
As in R1 to Rs, the cumulative number of appearances and the reading of characters are performed in parallel from the rank, and the arithmetic decoding unit 50
Is compared with the read cumulative appearance number of each rank, and in this case, the cumulative appearance number cf of the code between rank s and rank s-1
(Code) is located, the order s is restored, and T
The character output of order s is performed in 11 cycles.

As described above, in the T7 to T11 cycles,
Each time the ranks 1 to s are read, the numbers of adjacent cumulative occurrences in the rank higher by one are compared. In this case, when the comparison result at the time of reading the rank s-1 and the comparison result at the time of reading at the rank s are the same as the number of adjacent cumulative occurrences, the ranks s-2, s-1,. It can be seen that the cumulative number of occurrences of s is the same.

Therefore, in the cycles T11 to T13,
The stored data is replaced with s-2 at the highest position in the order in which the stored data of the restored character rank s has the same cumulative number of occurrences. Prior to this replacement, the cumulative number of occurrences of rank s−3 or higher read at T8 to T10 is incremented by one, and the total cumulative number of occurrences is read as R0 at T7 and then incremented by one at T8. And write at T9. Of course, T11-T
The replacement of the frequency data of the ranks s and s-2 in 13 is performed for three of the cumulative number of appearances, the number of appearances, and the characters.

The cycle from T14 in FIG. 15 to T2 in FIG.
In the case where the restoration order up to four cycles is the first place, the difference is that the frequency data in the restoration order in the probability model unit 52 is not replaced, and the rest is the same.
In the case where the restoration order is not the first place but the normal order in the cycles T25 to T38 in FIG. 16 and there is no higher rank having the same cumulative appearance number as the restored appearance cumulative number,
The difference is that the frequency data of the highest order and the restoration order having the same cumulative appearance number are not replaced, and the rest is the same as the normal case. 2. Partial Parallelization of Blend Context FIG. 17A shows a data compression apparatus according to the present invention for a blend context, in which a stochastic model part for blending is parallelized for each degree, and is further combined with an arithmetic coding part. It is characterized in that pipeline processing is performed between the two.

In this embodiment, the blending probability model unit 82 is constituted by a parallel circuit of a zero-order probability model unit 88 , a first-order probability model unit 90, and a second-order probability model unit 92.
Each of the zero-order, first-order, and second-order probability model units 88, 90, and 92 inputs one character at a time in parallel, obtains an appearance frequency that is dependent on each order, and outputs a rank. The buffer selection unit 84 includes a 0th-order, 1st-order, and 2nd-order probability model unit 8
The output order of 8, 90 and 92 is selected from the higher order to the lower order and output to the arithmetic encoding unit 86. Specifically, when the effective rank is obtained by the second-order probability model unit 92, this is output to the arithmetic coding unit 86 to perform arithmetic coding.

[0134] On the other hand, when the rank in the secondary probability modeling unit 92 becomes an escape output E to not obtained, and outputs a code simply indicates an escape without arithmetic coding, the following 1
The output of the next-order probability model unit 90 is selected, and if the effective rank is obtained, arithmetic coding of the order of the first-order probability model unit 90 is performed. When the effective order is not obtained from the first-order probability model unit 90 and the output is the escape output E, the output order of the zero-order probability model unit 88 is selected and arithmetic coding is performed. Here, the escape output in the second-order probability model unit 92 and the first-order probability model unit 90 is that when the input character is a parent and the characters of each order are children, there are children in each order, but there is no corresponding child Or no child of each order.

0th, 1st and 2nd order probability model units 88, 9
The appearance frequency data storage unit used in 0, 92 has a dictionary for every character to be encoded, and registers the character immediately before the input character used in the primary probability model unit 90 for each character input. And the second previous character used in the second-order probability model unit 92 is dynamically registered. FIG. 17 (B) shows FIG.
5A is a time chart of a pipeline operation of the data compression device in which the blending probability model unit 82 of FIG. Here, for example, in the order of the characters dcaba from the source data,
This is a case where aba is encoded after encoding of the character dc.

First, in the T1 cycle, the character a is input, and the 0th, 1st and 2nd order probability model units 88, 90, 9
In each of 2, the frequency of appearance in each order is obtained in parallel, and the order in the order of frequency of appearance is obtained. Here, the preceding two characters dc in the second probability model unit 92
Assume that the appearance frequency of the character string dca in which the character string dca is combined exists in the first place, and the buffer selection unit 8 in the next T2 cycle.
Reference numeral 4 selects the output 1 ₂ of the first rank output from the second-order probability model unit 92, and supplies it to the arithmetic coding unit 86, arithmetically codes it, and outputs a code 1.

At the same time, in the T2 cycle, the next character b is replaced by the 0th, 1st, and 2nd order probability model units 88, 90,
Are input in parallel with each other to determine the rank based on the appearance frequency.
At this time, if the second probability model unit 92 cannot retrieve the child of the preceding two characters ca, the arithmetic encoding unit 86 outputs the escape code E2 as it is in the next T3 cycle. In this T3 cycle, the blending probability model unit 82 inputs the third character a, and in this case, it is assumed that the secondary ranking model unit 92 has obtained the first place due to the presence of the preceding two characters ab. .

In the arithmetic encoding unit 86, following the encoded output of the escape output E2 from the secondary probability model unit 92 in the T3 cycle, the output of the primary probability model unit 90 is selected in the next T4 cycle. Try to encode. At this time, as shown in the cycle T2, the character a immediately before the input character b does not exist as a child in the first-order probability model unit 90, and the escape order E1 is obtained without obtaining an effective rank. Then, this escape output E
1 is output as code 1 as it is.

In the T5 cycle, the 0th-order probability model unit 8
The valid output 2 ₀ of rank 2 output from 8 encoded given to the arithmetic coding unit 86, and outputs a code 01. Therefore, 0 in the blending probability model unit 82 in the T2 cycle.
As for the rank output by the first-order and second-order parallel processing, two escape codes using the next three cycles of T3, T4, and T5 and the arithmetic coding of the zeroth-order rank output are performed.

Output 1 of rank 1 from the secondary probability model unit 92 in the T3 cycle of the blending probability model unit 82
Regarding ₂ , encoding is performed by the arithmetic encoding unit 86 in the next T6 cycle, and an output of code 1 is obtained. FIG. 18 shows the details of the data compression apparatus for the blend context shown in FIG. In FIG. 18, the blending probability model unit 82 includes a zero-order probability model unit 88 and a first-order probability model unit 9.
0, and a second-order probability model unit 92 in parallel.
It has a configuration shown as a representative of the first-order probability model unit 90.

The first-order probability model unit 90 includes a rank reading / updating unit 94, a character reading / updating unit 98, a frequency reading / updating unit 100, a cumulative frequency reading / updating unit 102, and a total cumulative frequency reading / updating unit 106.
Is provided. These blocks are the same as the corresponding circuit units provided in the probability model unit 14 for the fixed context in FIG. In addition, in the case of the blending probability model unit 82, an escape frequency readout update unit 95 and an escape determination unit 104 are newly provided.

The escape determination unit 104 detects whether the character immediately before the input character exists as a child in the corresponding dictionary of the frequency data storage unit 96, and if there is no child, escapes to the buffer unit 110. Force output.
The escape frequency reading / updating unit 95 increments the number of escape occurrences stored in the frequency data storage unit 96 by one when an escape is detected, and furthermore, the escape output is one of the number of character occurrences. Activate the total cumulative frequency reading / updating unit 106 to increment the total number of appearances.

The buffer selection section 84 is composed of buffer sections 108, 110, 112 provided for each order and a selection section 114. The selection unit 114 selects the order as the probability output stored in the order of the buffer units 112, 110, and 108 from the higher order to the lower order and supplies the order to the arithmetic encoding unit 86. The arithmetic coding unit 86 includes a code space calculation unit 116 for high-level calculation and a code space calculation unit 1 for low-level calculation.
18, a reciprocal calculating unit 120 for calculating the reciprocal of the total number of appearances, an arithmetic code deriving unit 122, and a code output unit 124. The configuration of the arithmetic coding unit 86 is the same as that of the corresponding block of the arithmetic coding unit 16 in FIG.

FIG. 19 is a flowchart of the pipeline operation of the primary model unit 90 and the arithmetic coding unit 86 in the data compression apparatus for the blend context shown in FIG. Here, when the T1 to T9 cycles are coded in other than the first place, the T10 to T13 cycles are escape coded when there is no child registration, and the T14 to T19 cycles have children but no corresponding children. Escape encoding in case.

In the operation of the first-order probability model 90, for example, when the cycle of T1 to T9 is taken as an example, T1, T1
In two cycles, the presence or absence of a child that is the character immediately preceding the input character is read, and if the order is obtained effectively, it is determined that there is a child, and the arithmetic coding unit 86 via the buffer selecting unit 84
The ranking is output to. The other points are the same as those of the fixed-order probability model section shown in the time chart of FIG. The arithmetic coding unit 86 as a common unit is exactly the same as the arithmetic coding unit 16 in FIG.

In the escape encoding in the case where there are no children in the T10 to T13 cycles, the first-order probability model 90
Only the arithmetic unit operates, and the arithmetic coding unit does not operate. Also T14
In the escape encoding in the case where there are children in the T19 to T19 cycles but there is no corresponding child, the escape output is encoded by the arithmetic encoding unit 86 by the same processing as in T1 to T8.

FIG. 20A shows data restoration for a blend context for restoring source data as a character string from compressed data decoded by a data compression device in which the blending probability model unit of FIG. 15 is parallelized. This is the basic configuration of the device. In FIG. 20 (A), the data restoration apparatus includes an arithmetic decoding unit 126 followed by a blending probability model unit 128.
And the blending probability model unit 128 is provided with, for example, a 0th-order probability model unit 130, a first-order probability model unit 132, and a second-order probability model unit 134 in parallel, and outputs the output of each model unit to a selection unit 136. Is selected to restore the string. Further, a pipeline control unit 125 is provided so that the arithmetic decoding unit 126 and the blending probability model unit 128 perform a pipeline operation.

The pipeline operation of the data restoration apparatus shown in FIG. 20A is as shown in FIG. Figure 1 now
It is assumed that the code 10101 1 obtained by the data compression device 7 is supplied. It is also assumed that the character dc has already been restored by the preceding two codes. First, in the T1 cycle, the code 1 is input to the arithmetic decoding unit 126, the second-order encoded output 1 ₂ is output, and the 0th, 1st, and 2nd order probabilities provided in the blending probability model unit 128 are provided. Model part 1
30, 132 and 134 are supplied in parallel. In the next T2 cycle, the first bit 0 of the second code is input, and the arithmetic decoding unit 126 outputs a secondary escape output E2. At this time In the probabilistic model 128, 2-order probability modeling unit 134 restores the character a the decoded output 1 ₂ of rank 1, and outputs via the selector 136.

In the next T3 cycle, bit 1 of the second code is input, and the primary escape output E1 is output by the arithmetic decoding unit 126. At the same time, the secondary probability model unit 134 updates the escape frequency data based on the secondary escape output E2 output in the cycle T2. In the T4 cycle, the bit 01 of the second code is input to the arithmetic decoding unit 126, and the 0-th order decoded output 2
Outputs ₀ . At this time, the blending probability model unit 128
In the case of (1), the primary probability model unit 132 updates the escape frequency data with the primary escape decoded output E1.

In the T5 cycle, the third code 1 is input, and the arithmetic decoding unit 126 outputs a _second- order first-order decoded output 12 . At this time, the blending probability model unit 1
In the 28 zero-order probability modeling unit 130 is operated from the decoded output 2 ₀ of the second rank to restore the character b output.
Decoded output 1 ₂ obtained by T5 cycle, restoration of character a by the secondary probability modeling unit 134 in cycle T6 is performed.

FIG. 21 shows details of the data restoration apparatus shown in FIG. The arithmetic decoding unit 126 includes a code input unit 138,
A cumulative value deriving unit 140, code space calculating units 142 and 144,
The reciprocal calculation unit 146, the cumulative frequency derivation unit 148, and the reciprocal calculation unit 150 are configured. The configuration and function of the arithmetic decoding unit 126 are the same as those of the arithmetic decoding unit 50 in the fixed-order data restoration device of FIG.

[0152] Blends probability model unit 128 includes a zero-order probability modeling unit 130, one Tsugi確ratio model unit 132 and a secondary probability modeling unit 134 has a configuration shown as a representative in the primary probability model 132 . The primary probability model unit 132 is provided with a character reading / updating unit 156, a frequency reading / updating unit 158, a cumulative frequency reading / updating unit 160, and a total cumulative frequency reading / updating unit 162. These blocks and functions are the same as those in the probability model unit in FIG. It is the same as 52 corresponding blocks.

Also, the order deriving unit 152 determines whether the order is 0, 1 or 2
It is provided as a common part to the next probability model units 130, 132, 134. Further, a frequency data storage unit 154 is provided as a common unit. Also, three selection units 136-1 and 1
36-2 and 136-3 are provided. Selector 136-1
Is selected in accordance with the order of the input code and the restored character of each order of the character reading / updating unit 156 is selected and output from the character output unit 164. The selection unit 136-2 receives the output of the cumulative frequency read / update unit 160 of each order, selects the output of the restored order, and selects the coding space calculation units 142 and 144 of the arithmetic decoding unit 126.
Output to The selection unit 136-3 receives the output of the total cumulative frequency read / update unit 162 of each order, selects the output corresponding to the restored order, and calculates the reciprocal of the selected order by the arithmetic decoding unit 126. Output to the unit 146.

FIGS. 22 and 23 are time charts of the pipeline operation of the data restoration apparatus shown in FIG. 21. The arithmetic decoding unit 126 as a common unit and the primary probability model unit in the blending probability model unit 128 are shown. 132 is shown as a representative. Further, the T1 to T13 cycles in FIG. 22 are for the case where decoding is performed in the first place restored from the input code,
The cycle from T14 in FIG. 22 to T24 in FIG. 23 is escape encoding when there is a child, and furthermore, T25 to T38.
The cycle is a normal case of encoding in a position other than the first place. 3. FIG. 24A shows another embodiment of a data compression apparatus for a blend context. In this embodiment, a probability model unit 166, an arithmetic encoding unit 168, a code Output unit 170
Are parallelized, and each part is further controlled by the pipeline control unit 1.
The pipeline operation is performed at 75.

That is, in the present embodiment, a second-order blend context is taken as an example.
6 is a zero-order, first-order, and second-order probability model units 172, 17
4,176, and the arithmetic coding unit 168 includes zero-order, first-order, and second-order arithmetic coding units 178, 180, and 182,
Further, the code output unit 170 includes zero-order, first-order, and second-order code buffer units 184, 186, and 188, and finally selects the code output of the order encoded by the merging unit 190 to obtain compressed data. .

FIG. 24B is a timing chart of the pipeline operation of the data compression device of FIG.
An example will be described in which a character dcaba is input and a character aba is encoded in a state where two characters dc have already been encoded. First, in the T1 cycle, the character a is changed to the 0th order, 1st.
Next, the order is input to the second-order probability models 172, 174, and 176 in parallel to determine the order.

At this time, the encoded output I ₂ of rank 1 by registering the encoded two characters dc in the second-order probability model unit 176
Is obtained. Here, when the higher-order output is effectively obtained, this is transmitted to the lower-order probability model unit to prohibit the output. In the cycle T1 Accordingly, only the secondary probability modeling unit 176 performs encoding output 1 ₂ of rank 1.

In the next T2 cycle, the next character b
Into the 0th, 1st and 2nd order probability model units 172, 174, 17
Enter 6 and obtain the ranking. At this time, the secondary arithmetic encoding unit 182 outputs the encoded output 1 ₂ from the secondary probability model unit 176.
And encodes the code 1 into the secondary code buffer 18.
Enter 8

In the T2 cycle of the probability model unit 166, the two-character ca
And produces an escape output E2. Also, even in the first-order probability model unit 174, the child character a is not registered and
The next escape output E1 is produced. When such an escape output occurs, the 0th-order probability model unit 172 becomes valid and generates a _0th- order encoded output 20 .

In the T3 cycle, the third character a is input, and the 0th, 1st and 2nd probability models 172, 174, 1
Parallel rank determined in 76, in this case, the secondary probability modeling unit 176 from the secondary encoded output 1 ₂ of rank 1 is obtained effectively. In this T3 cycle, 0 order, 1
Next, the secondary arithmetic coding units 178, 180, and 182 operate in parallel to store the codes 01, 1, 0 in the zero-order, first-order, and second-order code buffers 184, 186, 188.

[0161] In the next cycle T4, the encoded output 1 ₂ of rank 1 from the secondary probability modeling unit 176 obtained in T3 cycle is coded in the secondary arithmetic coding unit 182, the secondary code buffer unit 188 Is stored in FIG.
3A is a detail of a data compression device parallelized for each order. First, the blending probability model unit 166 determines the 0th, 1st,
Next and second order reading / updating sections 192, 194, and 196 are provided in parallel, and subsequently, 0-order, first-order, second-order, and cumulative frequency reading / updating sections 198, 200, and 202 are connected in series at the preceding stage to be in parallel. Is provided. Arithmetic encoding section 168 calculates the order 0, 1
Next, the code space calculation units 260, 262, 26 for each of the secondary
4. Arithmetic code deriving units 266, 268, 270, and serial / parallel conversion code output units 272, 274, 276
Is provided.

The code output of each order of the arithmetic coding unit 168 is the same as that of the code output unit 170 provided in the code output unit 170 of FIG.
Next, primary and secondary code buffer units 184, 186, 188
After that, they are compiled by the merge unit 190 and output as compressed data. The data structure shown in FIG. 25B is generated so that the codes of the respective orders can be restored independently at the time of collection by the merging unit 190.

The data structure generated by the merging unit 190 is as follows.
0th, 1st, and 2nd byte length display areas 266, 2 indicating the byte length of the 0th, 1st, and 2nd buffer codes at the beginning
68, 270, followed by a pipeline control unit 165 in which zero-order, first-order, and second-order buffer codes 260, 262, 264 are arranged, and a blending probability model unit 166, arithmetic coding The pipeline operation of the unit 168 is performed.

FIG. 26 is a time chart of the pipeline operation of the data compression device of FIG. First T1
25 shows the process of deriving the rank for each of the secondary, primary, and zeroth input characters and updating the frequency data in the blend probability model unit 166 of FIG. Also, the second order, 1
Although the next and zeroth-order arithmetic coding processes are shown in series in cycles T2 to T15 for convenience of space, the arithmetic coding processes are actually performed in parallel from the T2 cycle.

The pipeline operations of the probability model unit 166 and the arithmetic coding unit 168 in FIG. 26 are basically the same as those of the fixed order shown in FIG. FIG.
1 is a basic configuration of a data decompression device that decompresses compressed data encoded by the data compression device of FIG. This data restoration device includes a code input unit 212, an arithmetic decoding unit 214, and a probability model unit 216. The code input unit 212 is provided with a separating unit 211, for example, to input compressed data having the data structure of FIG.
Next, a code for each order can be stored in the secondary code buffer units 218, 220, and 222.

The arithmetic decoding unit 214 has 0-order, first-order, and second-order arithmetic decoding units 224, 226, and 228 provided in parallel. The blending probability model unit 216 includes a zero-order, a first-order,
Next probability model units 230, 232, and 234 are provided in parallel. Each restored character is input to the selection unit 236 and a character string is output. The pipeline operation of the data restoration device in FIG. 27A is as shown in FIG. This pipeline operation corresponds to code 1 compressed in FIG.
01011 is input to restore, and the operation is performed in a state where the character cd has already been restored. In the T1 cycle, the code 1 separated by the secondary code buffer unit 222 is output to the secondary arithmetic decoding unit 228, and the first encoded output 12 of the _{second order} is obtained.

In the next T2 cycle, the escape code 0 output from the secondary code buffer unit 222, the escape code 1 output from the primary code buffer unit 220, and the code output 01 from the 0th order code buffer 218 are input. Secondary, 1
Next, the zero-order arithmetic decoding unit 228,226,224 performs parallel decoding, secondary escape output E2,1 primary escape output E1, and produces a decoded output 2 ₀ of the zero-order rank 2.

[0168] In the this T2 cycle, and outputs the restored character a 2-order probability modeling unit 234 receives the decoded output 1 ₂ of first ranking obtained in T1 cycle simultaneously.
In the cycle T3, the output of the secondary code buffer unit 222 from the third code 1 is performed on the secondary arithmetic decoding unit 228 produces a coded output 1 ₂ of first ranking. At the same time, in the T3 cycle, the secondary probability model unit 234 updates the number of escape occurrences, and the primary probability model unit 232
, The number of escape occurrences is updated, and the 0th-order probability model unit 230 restores and outputs the second character b.

The decoded output I ₂ obtained by the secondary arithmetic coding unit 228 in the T3 cycle is restored and output by the second-order probability model unit 234 of the character a having the first rank in the T4 cycle. FIG. 28 shows details of the data restoration apparatus of FIG. 27A, and shows the arithmetic decoding unit 214 and the blending probability model unit 216. The arithmetic decoding unit 214 has 0
Next, primary and secondary parallel / serial conversion code input units 278, 280, 282, arithmetic code order derivation unit 28
4,286,288 and code space calculation units 290,29
2,294 are provided in parallel.

For this reason, the 0-order buffer code 260, the primary buffer code 262 and the secondary buffer code 264 separated by the code input unit 212 are input in parallel,
The cumulative appearance frequency for deriving the rank in the code space is calculated. The probability model unit 216 reads and updates the rank / frequency / cumulative frequency reading units 246, 248, and 2 for each of the 0th, 1st, and 2nd orders.
50 and character reading / updating units 252, 254, 256 are provided, and the character restored at any order is selected by the selection unit 258 and output. As described above, the pipeline operation is performed by the pipeline control unit 215 on the arithmetic decoding unit 214 and the blending probability model unit 216 provided in parallel for each order.

The time chart of FIG. 29 shows the pipeline operation of the data restoration device of FIG. First, regarding the bit code input, the cumulative appearance frequency calculation, and the coding space calculation in the arithmetic decoding unit 214, the 0th-order, first-order, and second-order processes are commonly shown. The next blending probability model unit 216 shows the parallel processing separately for the second order, first order, and zero order. The pipeline operation of the arithmetic decoding unit and the probability model unit in each order is basically the same as that of the arithmetic decoding unit and the probability model unit in the case of the fixed context shown in FIGS.

As described above, the arithmetic decoding section of each order and the probability model section operate independently, so that the speed of decoding processing can be improved. Even if the restoration process is executed halfway due to the escape code, it may become invalid. However, since the restoration process is independent for each order, there is no delay in the restoration process. 4. Access of Frequency Data Using Hash Structure FIG. 30 shows an embodiment of an access structure for reading and updating data in a frequency data storage unit used in a probability model unit of a data compression device and a data decompression device according to the present invention. Is characterized by using an internal hash structure.

In FIG. 30, the internal hash structure includes an internal hash function unit 300 and an appearance frequency model data storage unit 3.
02, a match / mismatch detection unit 304 and a rehash function unit 306. Appearance frequency model data storage unit 30
In the search of No. 2, a character string obtained by adding the immediately preceding character string having an arbitrary degree to the input character of interest is given to the internal hash function unit 300 to obtain a hash address, and the appearance frequency model data storage unit 302 is stored using the hash address. Search and
For example, a probability or a rank, which is an appearance frequency, is read.

The match / mismatch detection unit 304 determines whether or not the character string retrieved from the appearance frequency model data storage unit 302 matches the input character string based on the hash address. Output. In the case of a mismatch, since the hash function is registered at the rehashed address due to the collision of the hash address, the hash function unit 306 is activated to generate the hash function again, and again the appearance frequency model data storage unit 30
2 is searched, and the character string immediately before and the input character located thereunder are searched. Of course, the same applies to registration update.

FIG. 31 shows another embodiment of the appearance frequency data access structure provided in the appearance frequency model unit. In this embodiment, the immediately preceding character string (context) is searched, registered, and updated using an internal hash structure. The input character (attention character) existing under the immediately preceding character string (context) is searched, registered, and updated by a hash structure.
First, the internal hash structure includes an internal hash function unit 300,
It comprises an appearance frequency model data storage unit 302, a match / mismatch detection unit 304, and a rehash function unit 306. On the other hand, the external hash structure of the input target character is constituted by the target character data storage unit 308 and the match / mismatch detection unit 310.

For example, when an input character is set as a target character and the appearance frequency such as the probability or rank of the immediately preceding character string and the target character is read, the previous character string is given to the internal hash function unit 300 to obtain a hash address. The character string data is read out by accessing the appearance frequency model data storage unit 302. If the character string matches the immediately preceding character string, the attention character data storage unit 308 having an external hash structure is accessed, and the attention character located below the immediately preceding character string is matched / mismatched.
The list search is repeated until a match with 0 is obtained. Then, if the target character matches the character read from target character data storage unit 308 by the list search, the appearance frequency is output.

By performing such access using the internal hash structure shown in FIG. 30 or the combination of the internal hash structure and the external hash structure shown in FIG. 30, a context having an arbitrary degree in the appearance frequency model part in data compression and data restoration is obtained. The search, registration, and update of the appearance frequency of the dependent character can be performed at high speed. As for the compression / decompression for the blend context, arithmetic coding and decoding are taken as examples, but dynamic Huffman coding and decoding and spray coding and decoding which are a kind of self-organizing code are also used. Needless to say, the conversion may be performed. 5. Encoding and Decoding in Word Stream Direction FIG. 32 is a block diagram of a data compression device that encodes word data composed of a plurality of bytes in word units.
It is characterized in that a plurality of byte data constituting one word are encoded for each byte data in the byte stream direction.

In FIG. 32, when word data having an n-byte structure is taken as an example, the data compression apparatus has a byte array conversion section and n-byte byte unit coding sections 502-1 and 50-2.
2-2,..., 502-n, and a code merging unit 503. The byte array conversion unit 500 receives word data of one word, which becomes an n-byte byte stream from the source data, and separates and arranges each byte position in parallel.

That is, the first byte, the second byte,..., The n-th byte of the byte data are separated and arranged in the word stream direction for each byte position. The byte array conversion unit 500 stores, in addition to the currently input word data, byte data for each byte position of the preceding encoded word data.

As a result, the byte unit encoding unit 502-
For 1,502-2 to 502-n, each byte data at the same byte position of the current word data and the already encoded word data holds a byte data string in which the byte data is arranged in the word stream direction. . By holding the input byte data and the encoded byte data sequence at the same byte position for a plurality of words from the first byte to the n-th byte, the byte data sequence at each byte position becomes the byte data sequence in the word stream direction. Make up.

Byte unit encoding units 502-1 to 502-
Each of n performs dictionary-type coding or probability-statistic-type coding on a byte sequence in the word stream direction at each byte position. Dictionary coding includes LZW method and LZ method.
A typical method such as the SS method is used. As the coding of the probability statistical type, a decoder coding method such as multi-level arithmetic coding or dynamic Huffman coding or spray coding is used.

The code merging unit 503 collects the encoding of the byte data at each byte position of the input word data encoded in parallel by each of the byte unit encoding units 502-1 to 502-n and generates the compressed data. Output. The method of collecting the compressed byte data by the code merging unit 503 includes a method of uniformly arranging the compressed byte data at byte positions in the same byte format as the word data and collecting them in word units, and a method of combining the compressed byte data at each byte position into a file There are two methods of collecting compressed data in units.

FIG. 33 is a time chart of the encoding process in the word stream direction in the data compression apparatus shown in FIG. 32. The byte unit encoding sections 502-1, 502-2,
FIGS. 33A to 33C show the encoding process of the first byte, the second byte, and the n-th byte corresponding to 502-n. First, in the T1 cycle, each of the byte data from the first byte to the n-th byte of the first word is encoded by each of the byte unit encoding units 502-1 to 502-n. I have. in this case,
First, the encoding of the first byte is completed, then the encoding of the second byte is completed, and finally, the encoding of the nth byte is completed.

The encoded compressed byte data is stored in the buffer of the encoding merge unit 503 and is collected, and then output as compressed data at a predetermined timing.
The processing cycle T1 of the first word is the n-th byte unit encoding unit 502-n at the end of the encoding process.
This is the time when a certain time has elapsed from the processing end timing, and varies depending on the content of the encoding processing.

In the next T2 cycle, the parallel encoding of each byte data from the first byte to the n-th byte of the second word is performed by the byte unit encoding units 502-1 to 50-2.
2-n, and in this case, the encoding process of the second byte in FIG. 33B is the last. The next T3 cycle is the encoding of the third word, and the same processing is repeated until the last word data.

FIG. 34 is a block diagram of the code merging unit 5 shown in FIG.
13 is an encoding flowchart in a case where compressed byte data is grouped in words in 03. First, in step S1, n
Byte-structured word data is input, and in step S2,
The byte array conversion unit 500 divides the word data into bytes. Subsequently, in step S3, the byte unit encoding unit 502-
Encoding is performed in parallel in 1 to 502-n. Then, in step S4, each code data is combined in word units, and the processing in steps S1 to S4 is repeated until the encoding of the last word data is determined in step S5.

FIG. 35 is an explanatory diagram of the format of the compressed data in the case where the compressed byte data is grouped in words in FIG. The compressed data grouped in words is composed of a header 504 and compressed data 506. Header 504
Are provided with byte headers 504-1, 504-2,... 504-n for n bytes constituting one word data.

The byte headers 504-1 to 504-n
Each stores the number of compressed bytes at each byte position and the number of data at each byte position when the compressed byte data is put together in word units. Compressed data 504
Is a word having an n-byte configuration, like word data before encoding. At each byte position of the word data having the n-byte structure, the first byte compressed data, the second byte compressed data,..., the n-th byte compressed data are stored.

Since the compressed data of the corresponding byte data stored in each byte position has a different size of the compressed data, that is, the number of bytes, the compressed byte data of each byte data is uniformly stored by encoding in words. However, the actual number of compressed byte data at each byte position is different, and when overflow occurs, the data is stored at the next word position for storing compressed byte data.

FIG. 36 shows word data 508-1, 508-2, 50 by taking word data having a 4-byte structure as an example.
8 represents compressed data when four words of 8-3 and 508-4 are sequentially input and encoded. Word data 50
8-1 has the contents of A1, A2, A3, and A4 in byte order, which are compressed data 510-1 by encoding.
Are compressed into compressed byte data a1 to a4. The next word data 508-2 is byte data B1 to B1.
B4, and are compressed into compressed byte data b1 to b4 like compressed data 510-2 by encoding.

Word data 508-3 is byte data C
1 to C4, and compressed data 510-
3, compressed byte data c1 to c4.
Further, the word data 508-4 includes byte data D1 to D4.
The compressed data 510-4 is converted into compressed byte data d1 to d4 by encoding. Compressed data 51
Each of the compressed byte data in 0-1 to 510-4 has a different number of bytes. For example, the compressed byte data a1, a2, and a4 are 0.25 bytes, and the compressed byte data a3 is 0.5 bytes. ing. In the format of the compressed data to be summarized in word units in FIG. 35, the header 504 as shown in FIG.
The compressed data 506 as shown in FIG. 36 is put together from 510-1 to 510-4 obtained by encoding in word units in FIG.

First, the compressed data 510- in FIG. 36 for forming the compressed data 506 in word units in FIG.
The arrangement of 1 to 510-3 will be described. The compressed byte data a1, a2, a3, and a4 of the compressed data 510-1 obtained by encoding the first word data 508-1 are as follows:
The compressed word data 512-1 having the same 4-byte structure as the word data is evenly arranged at the corresponding byte positions from the first byte to the fourth byte.

That is, compressed byte data a1 is stored in the first byte of the compressed word data 512-1, compressed byte data a2 is stored in the second byte, and compressed byte data a3 is stored in the third byte. The fourth byte stores compressed byte data a4. Next word data 5
The compressed data 510 of FIG. 36 obtained by the encoding of 08-2
Similarly, compressed byte data b1, b2, b3, and b4 of -2 shown in FIG.
Each byte position of 12-1 is assigned as b1, b2, b3, b4. Similarly, compressed data 510-3 and 510-4 of the third and fourth words in FIG. 36 are also allocated as shown in FIG.

Here, the compressed byte data d1 of the fourth word
Among the compressed byte data d1, the first byte of the compressed word data 512-1 can be stored in the first byte, but the remaining compressed byte data d2 to d4 are two bytes of the first compressed word data 512-1. Since the first, third, and fourth bytes are full, they are stored in the corresponding byte positions of the next compressed word data 512-2.

When the compressed word data 512-1 and 512-2 in which the compressed byte data are stored in such four words are output collectively, the header 504 includes the first to fourth bytes of each compressed byte data. The number of bytes and the number of compressed byte data at each byte position in each of the compressed word data 512-1 and 512-2 are stored.

For example, in the byte header 504-1, the total number of the compressed data bytes a1 to d1 is one, and the first byte of the compressed word data 512-1 is the four compressed byte data a1, b1. , C1, and d1, the number of data 4 is shown. Since the number of data in the first byte of the next compressed word data 512-2 is 0, 0 is shown, and "4, 0" is set. .

By using the header information of the byte headers 504-1 and 504-4 for each byte position of the header 504, the compressed word data 512-1 and 512-2 arranged in word units can be used. 4 for each word
Compressed byte data corresponding to bytes can be separated. FIG. 38 is a flowchart of an encoding process when the encoded merged byte data is combined into a file unit at each byte position in the code merging unit 503 in FIG. First, in step S1, word data having an n-byte structure is input, and in step S2, a byte array conversion unit 500 is input.
Is divided into n byte data in byte units, and in step S3, the byte unit encoding units 502-1 to 502-n encode each byte data in parallel.

The parallel encoding in units of bytes in the word stream direction in steps S1 to S3 is repeated until the end of encoding of the last word data is determined in step S4. When the encoding of the series of word streams is completed, in step S5, the code data for each byte position, which are grouped in file units, are combined and output as compressed data.

FIG. 39 shows the format of the compressed data when the compressed byte data of FIG. 38 is put together in file units. The format of this compressed data is the header 514
And compressed data 516. The header 514 is provided with n file headers 514-1 to 514-n corresponding to each byte position of the word data. In the file header, the number of bytes n of the compressed data bytes of each file
1, n2,... Nn are stored. Compressed data 516
Stores compressed data 516-1 to 516-n in the first byte, the second byte,..., The nth byte in one file unit.

FIG. 40 shows an example of compressed data 510-1 to 510-4 obtained by encoding the 4-word word data 508-1 to 508-4 of FIG. It is a specific example of data. First, FIG.
In the compressed data 516 in the file unit of 0 (B), the byte files 516-1 to 516-51 are provided for each byte position of the first byte, the second byte, the third byte, and the fourth byte.
6-4 are provided.

Each byte file 516-1 to 516-4
Stores compressed byte data of the compressed byte data of 4 words by the encoding obtained in FIG. 36 corresponding to the byte position. According to the storage state of the compressed data 516 in file units, the header 5 in FIG.
14, each of the file headers 514-1 to 514-4 has a byte file 516-1 to 516-46.
4 stores the number of bytes of the compressed byte data and the number of data. The number of data is basically the same.

[0202] Even for such compressed data compiled in file units, the compressed byte data corresponding to the byte position of each word is separated from the code string at the time of decompression by using the header information of the header 514. be able to. FIG. 41 shows a specific embodiment of the data compression apparatus shown in FIG. 32, which is characterized by performing probability statistic coding on 4-byte word data.

In FIG. 41, first, the byte array conversion unit 5
00 is divided into the first byte, the second byte, the third byte, and the fourth byte of the word data, and a register group 518-1, 518-2, 518-3, and 5 having four registers is provided.
18-4 are provided. Here, the registers of the register group 518-1 at the first byte are R11 to R14, and the registers of the register group 518-2 at the second byte are R21 to R.
24, the registers of the register group 518-3 at the third byte are R31 to R34, and the register group 518 at the fourth byte
-4 are represented by R41 to R44.

The input source data is the first byte,
The first-stage registers R11, R21, and R31 are separated into second byte data, third byte, and fourth byte data.
And R41. Here, the word data is 32 bits, and since it is 4 bytes, it is 8-bit data per byte. Second stage register R
12, R22, R32, and R42 hold each byte data of the previously encoded word data.

The third-stage register group R13, R23, R3
3, R43 holds each byte data of the word data which has been encoded one immediately before. The last fourth-stage registers R14, R24, R34 and R44 have an additional 1
Each byte data of the previously encoded word data is held. As a result, in each of the register groups 518-1 to 518-4 of the first to fourth bytes, the byte data of the word data up to three words that have already been encoded from the current byte data are stored in each byte. The position is held in the word stream direction.

Since the byte unit coding unit 502 employs probability statistical coding, the appearance frequency model units 520-1 to 520-4 and the entropy coding units 522-1 to 522-4 are provided for each byte position. And the code buffer units 524-1 to 524-4 are provided at the last stage. Each of the appearance frequency model units 520-1 to 520-4 is
Using the correlation between the byte data to be currently encoded and the byte data of the past three words that have already been encoded, that is, the context, the ease of conditional appearance is detected by the appearance frequency model.

The appearance frequency model units 520-1 to 520-52
The entropy encoding units 522-1 to 522-4 perform statistical entropy encoding on the conditional ease of appearance (order) detected in each of 0 to 4, and are compressed data in byte units. The code is converted into a code and temporarily stored in each of the code buffer units 524-1 to 524-4.

In each of the code buffer units 524-1 to 524-4, when the compressed byte data is accumulated in units of bytes according to the format of FIG. 35 or in units of files according to the format of FIG. The code at each byte position is combined by the merging unit 533 to create and output a series of compressed data together with the header. FIG.
42 is a time chart of an encoding process for each byte of the data compression device that performs the probability statistical encoding shown in FIG. That is, regarding the appearance frequency model processing and the entropy encoding processing, FIG. 42A shows the first byte, FIG. 42B shows the second byte, FIG. 42C shows the third byte, and 42 (D) shows the fourth byte. Indicates the byte. For example, in the T1 cycle, the appearance frequency model processing of the byte data at each byte position is performed in parallel.

In this example, the appearance frequency model processing is completed in the order of the third byte, the second byte, the first byte, and the fourth byte. When the appearance frequency model processing ends, the entropy encoding processing of the T2 cycle starts a predetermined time after the last end timing, and at the same time, performs pipeline processing for executing the appearance frequency model processing in parallel for the next word data. I have. Hereinafter, similarly, T3, T4,
The appearance frequency model processing and entropy encoding processing of each cycle of T5 and T6 are performed in parallel for each byte position.

FIG. 43 is a basic block diagram of a data decompression device that inputs compressed data output from a data compression device that performs byte-by-byte encoding in the word stream direction in FIG. 32 and decompresses the data into word data units. It is. This data restoration device includes a code separation unit 530, byte unit decoding units 532-1, 532-2, ... 532-n corresponding to respective byte positions of n-byte word data to be restored, and an output switching unit 533. It consists of.

The code separation unit 530 inputs compressed data in accordance with the format having unity in byte units shown in FIG. 35 or the format having unity in file unit in FIG. 39, and converts each byte position into word data units based on header information. The code which is the byte compressed data of
The data is output in parallel to the byte unit encoding units 532-1 to 532-n.

Byte unit decoding units 532-1 to 532-
Each of n inputs the code of each byte position, and restores each byte data from the 1st byte to the nth byte in parallel by the reverse operation of the encoding. The switching output unit 533 combines the byte data output from the byte unit decoding units 532-1 to 532-n in the order of the byte positions, restores n-byte word data, and outputs it as source data.

Byte unit decoding units 532-1 to 532-
As n, dictionary type decoding or probability statistical type decoding is performed in correspondence with the data compression apparatus of FIG. As the dictionary-type decoding, a method represented by the LZW method or the LZSS method is used. As the probability statistical decoding method, a multi-level arithmetic coding method, a dynamic Huffman coding method, or a code tree coding method such as a spray coding method can be used.

FIG. 44 shows the decoding process in byte units in the data restoration apparatus of FIG. 43, and FIGS. 44 (A) and 44 (B)
(C) shows each decoding process of the first byte, the second byte, and the nth byte. For example, in the T1 cycle, the code at each byte position consisting of one byte of compressed byte data separated by the code separation unit 530 is input in parallel to the byte unit coding units 532-1 to 532-n, Each byte data is restored after different encoding processing times.

In the T1 cycle, the decoding process of the n-th byte is the last, and thereafter, from the timing when a predetermined time has elapsed, the sign of the code at each byte position corresponding to the next word data in the next T2 cycle Parallel encoding is performed and this is repeated. FIG. 43 is a flowchart of a process of restoring compressed data compiled in words in accordance with the format of FIG. First, in step S1, compressed data is input. In step S2, a code corresponding to each byte position of n bytes of word data is separated from the header information of the compressed data, that is, the byte compressed data is separated into one byte unit. The separated byte compressed data is stored in step S
In step 3, decoding is performed in parallel.

If it is determined in step S4 that the compressed data is not insufficient, the flow advances to step S5 to combine the decoded decompressed byte data in byte units and output one word of word data. The processes in steps S3 to S5 are repeated until the restoration of the final data is completed in step S6. FIG. 46 is a flowchart of a process for restoring compressed data compiled in file units shown in the format of FIG. First, in step S1, the compressed data is input collectively,
In step S2, the code included in the compressed data, that is, each byte compressed data, is all separated for each byte position of the word data to be restored. Subsequently, in step S3, decoding is performed in parallel on each of the separated compressed data in word units.

At this time, if there is no shortage in the compressed data, the flow advances to step S5 to combine the decompressed byte data of one word and output word data. Steps S3 to S
Step 5 is repeated until the restoration of the last word data is completed in step S6. FIG. 47 shows a specific embodiment of the data restoration apparatus of FIG.
This is a data decompression device corresponding to a data compression device that performs probability statistical coding in a byte configuration.

The data decompression device shown in FIG.
A byte unit decoding unit 532 is provided following 30.
Since one word is composed of 4 bytes, the first byte,
The code buffer units 534-1 to 534-4, the entropy decoding units 536-1 to 536-4, and the appearance frequency model units 538-1 to 538-4 correspond to the second byte, the third byte, and the fourth byte, respectively. 538-4 are provided in parallel.

Each of the code buffer units 534-1 to 534-4 temporarily buffers the code at each byte position which becomes one byte of compressed byte data separated by the code separation unit 530. Entropy decoding units 536-1 to 53-53
Step 6-4 detects the conditional ease of output from the input code and the byte data string immediately before being restored. The appearance frequency model units 538-1 to 538-4 register, as frequency data, the conditional ease of currently restored byte data connected to the immediately preceding restored byte data string as frequency data, and the entropy decoding unit 536 -1 to 536-4
The frequency data is retrieved by the conditional ease determined in each step to restore the byte data, and the frequency data is registered and updated.

Appearance frequency model units 538-1 to 538-4
The output switching unit 533 is provided with a group of registers 540-1 to 540-4 for storing three words of byte data restored for each byte position. Here, the register group 540-1 of the first byte is the register R11
0-R130, second byte register group 540
-2 is indicated by registers R210 to R230.

The register group 540-3 at the third byte is indicated by registers R310 to R330, and the register group 540-4 at the fourth byte is indicated by registers R410 to R430. Such a register group 540-1 to 540
-4, the appearance frequency model units 538-1 to 538- of the three bytes of the restored byte data of each byte position at each byte position
No. 4 is registered and updated in order to obtain the conditional ease of byte data to be restored and taken into connection with the immediately preceding byte data string.

The output switching unit 533 extracts the restored byte data at each byte position held in the registers R110, R210, R310, and R410 as source data of the first to fourth bytes. Are connected in byte order and output as word data. 48 is a time chart of the entropy decoding process and the appearance frequency model process in the data restoration device of FIG. 47. FIG. 48A shows the first byte, FIG. 48B shows the second byte, and FIG. The third byte is added to C), and FIG.
(D) shows the fourth byte.

For example, taking the T1 cycle as an example, for each of the first to fourth bytes, the code of each separated byte position is input, and the entropy for obtaining conditional ease from the immediately preceding restored byte data string is obtained. A decoding process is performed. In this case, for example, a predetermined time after the timing at which the entropy decoding process of the fourth byte is finally completed, a two-stage appearance frequency model process is performed as a T2 cycle.

At the same time, the code of each separated byte position for restoring the next word data is input, and the entropy decoding processing is performed in parallel. Hereinafter, T3, T4,
Similar processing is repeated as in a T5, T6... Cycle. FIG. 49 shows the appearance frequency model unit 5 provided in each of the first to fourth bytes of the data compression apparatus shown in FIG.
This is an embodiment of pipeline control of 20-1 and the entropy coding unit 522-1, and exemplifies a case where calculation coding is applied as probability statistical coding.

In this arithmetic coding, as in the pipeline control of FIG. 6, the probability model unit 14 and the arithmetic coding unit 1
6, and a pipeline control unit 15 is provided for this. In the arithmetic coding pipeline control of FIG. 49, the probability model unit 14 is divided into a probability model search unit 542 and a probability model update unit 544, and the arithmetic coding unit 16
Are divided into a code space operation unit 546 and an arithmetic code output unit 548.

Here, the probability model section 14 has the configuration shown in FIG. 7, and the probability model search section 542 corresponds to FIG.
A probability model search unit 542 that combines the functions of the order reading / updating unit 18, the character reading / updating unit 20, the frequency reading / updating unit 22, the cumulative frequency reading / updating unit 24, and the total cumulative frequency reading / updating unit 26 as a reading unit. On the other hand, the probability model updating unit 544 summarizes the functions of the remaining updating units.

On the other hand, the arithmetic encoding unit 16 is divided into a code space operation unit 546 and an arithmetic code output unit 548. The code space calculation unit 546 has functions of the code space calculation unit 28, the code space calculation unit 30, and the reciprocal calculation unit 32 in the arithmetic coding unit 16 in FIG. The arithmetic code output part 54
8 has the functions of the arithmetic code deriving unit 34 and the code output unit 36 of the arithmetic coding unit 16 in FIG.

The pipeline control unit 15 differs from the embodiment of FIGS. 6 and 7 in that the processing of the probability model update unit 544 and the code space calculation unit 546 is pipelined. FIG.
9 (B) is a pipeline process of the arithmetic coding in FIG. 49 (A), and, as in FIG. 6 (B), exemplifies coding when a character aba is input. 6B is different from FIG. 6B in that a pipeline process is performed in which two processes of a probability model update unit 544 and a code space calculation unit 546 are performed simultaneously in parallel after the probability model search unit 542. As a result, the processing performance by pipelining is further improved as compared with the cases of FIGS.

FIG. 50 shows an embodiment in which arithmetic decoding is applied to the entropy encoding process and the appearance frequency model process performed for each byte position from the 1st byte to the 4th byte of the data restoration device of FIG. is there. Of course, FIG. 50 shows one of the 4-byte parallel processing units as a representative. When arithmetic decoding is applied to the data restoration device, the entropy decoding process and the appearance frequency model process include an arithmetic decoding unit 50 and a probability model unit 52. The configuration of the arithmetic decoding unit 50 is the same as that of FIG. Details of FIG. 13A are shown in FIG.

The arithmetic coding unit 50 shown in FIG. 50 is divided into an arithmetic coding input unit 550, a cumulative value calculation unit 552, and a code space calculation unit 554. Looking at the correspondence with the embodiment of FIG. 14, the arithmetic coding unit 550 corresponds to the code input unit 54 of FIG. The cumulative value calculating unit 552 corresponds to the cumulative value deriving unit 56, the cumulative frequency deriving unit 64, and the reciprocal calculating unit 66. The code space calculation unit 554 includes a code space calculation unit 58,
60 and the reciprocal calculation unit 62.

The probability model section 52 in FIG. 50 is divided into a probability model search section 556, an output character search section 558, and a probability model update section 560. Probability model part 5 in FIG.
Looking at the correspondence with No. 2, the probability model search unit 556 and the output character search unit 558 are the order derivation unit 68, the character read update unit 72, the frequency read update unit 74, and the cumulative frequency read update unit 7 of FIG.
6 and a function as a reading unit in the total cumulative frequency reading / updating unit 78, and further includes a character output unit 80. The probability model updating unit 560 is a collection of functions as each updating unit in FIG.

In the arithmetic coding unit 50 and the probability model unit 52 in FIG. 14, the pipeline processing is performed in parallel with each other, but in the embodiment in FIG. Code space operation unit 554 and probability model unit 52
The difference is that the probability model search unit 556 also performs pipeline processing by parallelization. FIG. 50 (B)
FIG. 50A is a decoding process in FIG. 50A, in which a character is restored for the same code input as in FIG. 13B. The output of the probability model search unit 556 is output from the output character search unit 558.
And the code space operation unit 554 are supplied in parallel to enable pipeline operations at the same timing, thereby improving processing performance.

FIG. 51 shows an embodiment in which code tree coding such as dynamic Huffman coding or splay coding is applied to the appearance frequency model portion and entropy coding processing at each byte position in FIG. The embodiment of the code tree coding includes a context collection processing unit 562 and a code tree coding unit 564.
It consists of. The context collection processing unit 562 includes a context tree search unit 568, a context tree registration unit 570, and a context determination unit 572, and collects the context of the input byte data to create a context tree. The code tree coding unit 564 includes a code tree search unit 574, a code tree update registration unit 570, and a code rearrangement output unit 578, and performs spraying in accordance with a result of context grinding of input byte data by the context collection processing unit 562. Create and update a code tree while performing coding or dynamic Huffman coding.

In particular, in this embodiment, the code tree encoding section 564 makes the code tree search and registration section 574 parallel to the code tree update registration section 576 and the code rearrangement output section 578.
51, the search result of the code tree search unit 564 is input in parallel to the code tree update registration unit 576 and the code rearrangement output unit 578, as in the example of the encoding of the character string aba in FIG. In addition, the pipeline operation is performed at the same timing to shorten the encoding processing time.

FIG. 52 shows an embodiment for code tree decoding corresponding to the code tree coding shown in FIG. 51, and comprises a code tree decoding section 580 and a context collection processing section 582. The code tree restoration unit 580 includes a code tree search unit 586 and a code tree update registration unit 588. The code tree restoration unit 580 receives code data and performs code tree decoding such as spray decoding or dynamic Huffman decoding. Is created and updated, and the context of the restored byte data is collected to create a context tree.

In this embodiment of the code tree decoding,
The search result of the code tree search unit 586 is stored in the code tree update registration unit 58
52 is input to the context tree search unit 590.
As shown in (B), the processing of the code tree update registration unit 588 and the context tree search unit 590 is performed in parallel on the search result of the code tree search unit 582 at the same timing to form a pipeline, thereby shortening the restoration processing time. ing.

The details of the case where the splay coding (Splay-Tree) is used for the coding tree coding in FIGS. 51 and 52 are shown in Japanese Patent Application Laid-Open No. 8-30432 (published on Feb. 2, 1996). It is. 6. FIG. 53 shows an adaptive switching process between the word stream direction and the byte stream direction according to the present invention shown in FIG. 32. 1 is an embodiment of a data compression device that switches the data in a dynamic manner. The data restoration apparatus that encodes the byte stream and the word stream in parallel includes a byte array conversion unit 600, a word stream encoding unit 602, and a byte stream encoding unit 60.
4 and a code switching unit 605.

The byte array conversion unit 600 inputs word data having a plurality of bytes in units of words as source data, separates the data into byte data at each byte position, and sends the byte data to the word stream encoding unit 602. The position is input as a data byte in the word stream direction in which a plurality of byte data are arranged in the word direction at the position, while the byte stream direction having the data byte sequence of the data stream itself of the word data is input to the byte stream encoding unit 604. Input as byte data.

The byte array section 600 can be realized by the same register group as the byte array conversion section 500 shown in FIG. 41 taking, for example, a 1-word 4-byte configuration. The register arrangement direction of the same byte position in the register group of FIG. 41 is the word stream direction, and the vertical direction that is each byte position orthogonal to this is the byte stream direction. The encoding in the word stream encoding unit 602 has the basic configuration shown in FIG. 32, and more specifically, the one-byte 4-byte configuration shown in FIG. 41 can be used. Further, a specific example of the probability statistical type encoding at each byte position For example, the arithmetic coding shown in FIG. 42 or the code tree coding shown in FIG. 50 can be used.

On the other hand, byte stream coding section 60
4 is basically composed of a word stream encoding unit 602.
However, the difference is that the arrangement of byte data for one word input from the byte array conversion unit 600 is in the byte stream direction. The code switching unit 605 includes:
Based on the results of the parallel encoding performed by the word stream encoding unit 602 and the byte stream encoding unit 604, one of them is selected and output as compressed data.

FIG. 54 is a time chart of the word stream and byte stream encoding switching processing in FIG. 53. FIG. 54 (A) shows the word stream encoding processing, and FIG. 54 (B) shows the byte stream encoding processing. FIG. 54 (C) shows the selection code output. For example, in the first T1 cycle, the word stream encoding unit 6
02 and the byte stream coding unit 604 perform parallel coding on each byte data obtained by separating the word data in the byte array conversion unit 600, and perform dictionary coding or probability statistical coding in coding. The difference is whether the encoded word data is in the word stream direction or the byte stream direction.

In the T1 cycle, the byte stream encoding shown in FIG.
The word stream encoding of (A) has been completed. When the word stream encoding is completed, the next T
The word stream encoding and the byte stream encoding of the word data in two cycles are started. In parallel with this, the selection code of the word stream code or the byte stream code shown in FIG. Is output.

That is, in the encoding of the T1 cycle,
For example, the byte stream code 608 is selected, and the selection code 606 indicating the selection of the byte stream code 608 is selected.
Code output is performed by adding -1 to the beginning. In the next T2 cycle, the word stream code 610 is selected. Similarly, in the T3 cycle, first, the selection code 606-2 indicating the selection of the word stream code 610.
, The word stream code 610 is output as compressed data.

The selection of the word stream code and the byte stream code by the code switching unit 606 in FIG. 53 may be either byte-based switching or word-based switching. FIG. 55A shows a format of code switching in byte units. For example, taking the encoding of word data having a structure of one word and four bytes, in each of the word stream encoding unit 602 and the byte stream encoding unit 604, the first byte, the second byte, the third byte,
Each byte data of the byte is encoded in parallel.

Therefore, when both encodings are completed, 1
The byte stream code of the byte is compared with the byte stream code, and if, for example, the byte stream code is selected, one byte compressed data 608 composed of the byte stream code is output following the switching code 606-1. Next, the word stream code of the second byte is compared with the byte stream code. If, for example, the word stream code is selected, the switching code 606-2 is added to output the word stream code as one byte of compressed data 610. I do. In the same manner, the addition of the switching code in byte units and the output of 1-byte compressed data are repeated in the same manner.

FIG. 55B shows code switching in word units. When encoding of each word data for one word is completed in parallel, compressed data in word units, that is, compression for the number of bytes of one word, is performed. If the byte data is compared and, for example, the byte stream code side is selected, a byte stream code of compressed byte data for n bytes constituting one word is output following the switching code 612.

FIG. 56 is a flowchart of an encoding process in the case of performing code switching in byte units in FIG. 55 (A).
First, in step S1, word data having an n-byte structure is input, and in step S2, the data is divided into byte units.
In step 3 and step S4, byte-by-byte encoding in the word stream direction and byte-by-byte encoding in the byte stream direction are performed in parallel.

Subsequently, in step S5, the encoding efficiency in the word stream direction and the byte stream direction is compared for each byte, and the code with the higher efficiency is selected for each byte in step S6, and a switching or selection code is added. In step S7, the code data for each byte is combined and output. Such processing of steps S1 to S7 is repeated until the encoding of the final word data is completed.

FIG. 57 is a flowchart of the encoding process in the case where code switching is performed in word units in FIG. 55 (B). First, in step S1, word data having an n-byte structure is input, and in step S2, the word data is divided into n-byte units. In steps S3 and S4,
The encoding in the word stream direction and the encoding in the byte stream direction for each byte data are performed in parallel.

When the encoding is completed, the encoding efficiency in the word stream direction and the encoding efficiency in the byte stream direction are compared for each word in step S5. Then, in step S6, a code in the stream direction with good code efficiency is selected for each word from the comparison result, and switching or addition of a selected code is performed. Such processing of steps S1 to S6 is repeated until the encoding of the final word data is completed in step S7, and finally, in step S8, the respective encoded data are combined and output as compressed data.

FIG. 58 shows an embodiment in which one word of the data compression apparatus for performing code switching by performing coding in the word stream direction and the byte stream direction in FIG. 53 in parallel is constituted by 4 bytes. The case where probability statistical coding is applied as an example is described. FIG. 58 shows an embodiment of the byte array conversion unit 600, the word stream coding unit 602, and the byte stream coding unit 604 of FIG. In the byte array conversion unit 600, register groups 618-1 to 618-4 in which four registers are connected in series for the first byte, the second byte, the third byte, and the fourth byte, respectively, as a register group for context capture. Provided.

If the 4-byte source data input for the encoding process is, for example, 32 bits, the source data is separated into 8 bits per byte, and is divided into first-stage registers R1, R2, R3, and R4 in parallel. Is held. At this time, the second-stage registers R1 to R4 and the third-stage register R1
The third to fourth registers R14 to R44 hold the byte data at the respective byte positions of the preceding three words that have already been encoded.

With regard to such a register group for context capture, for example, when focusing on the register R41 storing the byte data of the fourth byte of the word data to be encoded, the byte data in the word stream direction is R41 to R44. In a row. On the other hand, the byte stream direction is the direction of R31, R32, and R11 in which the preceding byte data is sequentially stored in the register R11.

For the byte stream data stored in the group of registers provided in the byte array conversion unit 600, the word stream coding unit 602 sends the data at each byte position of the registers R11 to R14 to be currently coded. , Three bytes preceding the word stream in the word stream direction are input for context capture. That is, the word stream encoding unit 600 converts the source data 1
Probability model units 620-1 to 620-4, entropy coding units 622-1 to 622-4, and buffer selection units 624-1 to 62-4 corresponding to the second, third, and fourth bytes, respectively. 624-4. Each of the probability model units 620-1 to 620-4 is a 0th-order, a 1st-order,
Four probabilistic model units of second and third order are provided in parallel.

In response, entropy coding section 62
As for 2-1 to 622-4, 0th order, 1st order, 2nd order, 3rd order
The following entropy encoding units are provided in parallel. That is, the word stream encoding unit 600 uses a data compression device in which all the blend contexts shown in FIG. 24 are parallelized. Details of the data compression apparatus that is fully parallelized with respect to the blend context of FIG. 24 are described in FIGS.
Is the same as that shown in FIG.

The input from the context fetch register of the byte array conversion unit 600 to the probability models 620-1 to 620-4 provided at the first stage of the word stream coding unit 600 is as shown by the register numbers R11 to R44. is there. For example, the first byte probability model unit 620-1
Inputs the byte data of the first byte of the register R11 to be encoded to the zero-order probability model unit.

The first-order probability model unit receives the byte data of the register R11 of the first byte which has been previously encoded, and the second-order probability model unit receives the immediately preceding encoding. The byte data of the register R13 of the completed first byte is input, and another one is sent to the third-order probability model unit.
The byte data of the register R14 of the first byte that has been encoded immediately before is input. This point is based on the model part 620-2 to 620-4 of the second byte, the third byte, and the fourth byte.
For the register at each byte position.
Next, primary, secondary, and tertiary register inputs are performed.

On the other hand, byte stream encoding section 604
Are the first-stage probability model units 630-1 to 630-4, the next-stage entropy encoding units 632-1 to 632-4, like the word stream encoding unit 602 that realizes all parallelization for the blend context. And the buffer selection unit 634
1 to 634-4 are provided in parallel for each byte position.

The input from the context fetch register group of the byte array conversion unit 600 to the byte stream encoding unit 620 is the probability model units 630-1 to 630-4.
Starting from R11, R21, R31 and R41 for the 0th-order probability model unit for inputting byte data at each byte position of the word data to be currently processed, register input of byte data having a sequence in the byte stream direction is performed. Is going.

That is, the probability model part 630-at the fourth byte
In the case of No. 4, the fourth byte of the currently processed word data is input to the 0th-order probability model section, the preceding 3rd byte of the same word data is input to the first-order probability model section, and The register R21 of the second preceding word data of the second byte is input to the second-order probability model unit,
Register R11 of the first byte of the first byte
Is input to the third-order probability model unit.

The register R3 holding the third byte data of the next word data to be currently processed.
In response to the input from the 0th-order probability model unit to the probability model unit 630-3 from 1 the second byte register R21 of the same word data one byte before the first-order probability model unit
Is input to the second-order probability model unit, and the first byte register R11 which is one immediately before the same word data
Byte data is input to the third-order probability model part.
The byte data of the register R42 of the fourth byte before the word is input.

The byte data of the second byte is stored in the register R2.
Probability model unit 630 input to 1st to 0th order probability model unit
-2, R11, R4 in the byte stream direction.
2, R32 are input to the primary, secondary, and tertiary probability model units, respectively. Further, also regarding the probability model unit 630-1 in which the byte data of the first byte is input from the register R11 to the 0th-order probability model unit, each byte data of the registers R42, R32, R22 preceding in the byte direction is converted to the primary,
It is input to the second and third order probability models. Words are not considered for these byte-wise inputs in the byte stream direction.

The word stream coding section 602 shown in FIG.
And by the byte stream encoding unit 604 performing parallel encoding for each byte position,
4-1 to 624-4 and buffer selection units 634-1 to 634-6
34-4, the code W1 is used as the word stream code.
To W4 and the bit numbers BW1 to BW4 are output, and the byte stream codes B1 to B4 and the respective bit numbers BB
1 to BB4 are output.

FIG. 59 shows an embodiment of the code switching unit 605 shown in FIG. 53. The word stream coding unit 602 shown in FIG.
And the respective encoded outputs of the byte stream encoding unit 604 are input in parallel. The code switching unit 605 is provided with code switching units 648-1 to 648-4 and code buffers 650-1 to 650-4 in parallel for the first byte, the second byte, the third byte, and the fourth byte. Finally, the code merge unit 652 outputs the compressed data as a whole.

Code switching section 648-1 for each byte position
648-4 to 648-4, the number of bits comparison units 640-1 to 640-4, the word / byte selection probability model units 642-1 to 642-4, and the encoding unit 6
46-1 to 646-4. Bit number comparison unit 6
40-1 to 64-4 are the word stream codes W1 to W4 restored for each bit position and the bit numbers BW1 to BW4 and BB1 to B4 of the byte stream codes B1 to B4.
B4 are compared with each other, the shorter code length is selected, and the selection result is expressed as a word / byte selection probability model unit 642-1 to 64-2.
Output to 2-4. Word / byte selection probability model part 6
42-1 to 642-4 calculate probabilities from the frequencies of word stream codes and byte stream codes.

Encoding sections 646-1 to 646-4 generate code switching codes from word / byte selection probability model sections 642-1 to 642-4, and code switching section 648-1.
To 648-4, and at the same time, a selection code is added to the beginning. Code buffers 650-1 to 650
In the case of -4, when a certain amount of code data is accumulated in word units or file units, the data is combined again by the code merge unit 652 and output as a series of compressed data.

FIG. 60 shows an embodiment of a data decompression device for decompressing original word data from compressed data obtained by switching and selecting after encoding in the word stream direction and the byte stream direction in FIG. This data restoration device includes a code separation unit 700, a word stream decoding unit 702, a byte stream decoding unit 704, and an output switching unit 706.

The code separation unit 700 is either a compressed data of a format having a switching code in units of compressed byte data in FIG. 55 (A) or a format having a switching code for each compressed byte data of words in FIG. 55 (B). The compressed data is input, divided into a word stream code and a byte stream code based on the switching code included in the compressed data, and input to the word stream decoding unit 702 or the byte stream decoding unit 704.

The word stream decoding unit 702 operates as shown in FIG.
2. Basically the same as the data restoring device shown in FIG. 41, the code for the number of bytes constituting one word is decoded in parallel, and the byte data at each corresponding byte position is restored. The byte stream decoding unit 704 has a decoding function in parallel with the decoding function for each byte position forming one word in the same manner as the word stream decoding unit 702. And the byte unit decoding units 502-1 to 502-n are provided in parallel.

For example, taking the probability statistical type coding as an example, the same circuit as the byte unit decoding unit 532 in FIG. 47 can be used. The only difference between the byte stream decoding unit 704 and the word stream decoding unit 702 is that the byte stream decoding unit 704 receives and decodes the byte stream code separated by the code separating unit 700. FIG. 61 is a time chart of the decoding process in the data restoration apparatus of FIG. 60. FIG. 61 (A) shows the selection code decoding, FIG. 61 (B) shows the word stream decoding, and FIG. Represents byte stream decoding. First, in the first T1 cycle, the code separation unit 700 decodes the switching selection code provided at the head of the input compressed data, and determines whether the next compressed data is a word stream code or a byte stream code. Identify.

In the T1 cycle, the decoding of the selected code is identified as a word stream code,
The code data is supplied to the word stream decoding unit 702, and the byte data is decoded. Assuming that the byte stream code is determined by decoding the selection code 708-2 in the next T2 cycle, the code data is provided to the byte stream decoding unit 704, and the byte data is decoded. Hereinafter, similarly, selection codes 708-3, 70
According to the identification of the word stream code or the byte stream code according to the decoding result of 8-4,..., The corresponding decoding is performed to restore the byte data.

FIG. 62 is a flow chart showing a data decompression process for compressed data in a format in which a switching code is added to each code which is byte-based compressed data shown in FIG. It is. First, each code data is input-separated in step S1, and if it is not final data in step S2, 1
Check whether the code for the word could be input separated,
If the code for the word has been input-separated, step S4
Decodes each selected code.

On the basis of the decoding result of the selected code, in step S5, it is determined whether the code for the number of bytes constituting one word is a word stream code or a byte stream code.
The word stream decoding unit 7 of FIG.
02 or the byte stream decoding unit 704. Subsequently, in steps S6 and S7, each code is decoded in parallel for each byte in the word stream direction and the byte stream direction in parallel.

Subsequently, in step S8, the restored byte data in the word stream direction and the restored byte data in the byte stream direction restored for each byte are stored in the buffer. Then, in step S9, when the byte data for one word is prepared, in step S10, the restored byte data is combined in byte order and one word data is output. Subsequently, the process returns to step S2 to check whether or not the restoration of the final data has been completed. If the restoration has not been completed, the process returns from step S3 to step S1 again to repeat the same processing.

The flowchart of FIG. 63 is a data decompression process for compressed data in a format in which a selection code is added in units of words of FIG. 55B. First, in step S1, all the compressed data to be restored are collectively input and separated into codes. Subsequently, if the restoration of the final data has not been completed in step S2, the process proceeds to step S3, where the first selected code separated from the input is decoded, and a word stream code or a byte stream code is identified.

If it is a word stream code, step S
Proceeding to 5, the codes for the number of bytes constituting one word are decoded in parallel to restore each byte data. If the code is a byte stream code, the process proceeds to step S6, where 1
The code for the number of bytes constituting the word is decoded in parallel to restore each byte data. Next, in step S7, the decoded data in the word stream direction or the byte stream direction decoded for each word is stored in the buffer. Then, in step S8, the restored word data having an n-byte structure is output, and this is repeated until the last word data is restored in step S2.

Next, as a specific embodiment of the data compression apparatus of FIG. 60, the data compression apparatus of FIG.
A specific embodiment of a data decompression device for decompressing compressed data obtained by a data compression device in which all of the blend contexts shown in FIG. 9 are parallelized will be described. In a data decompression device for decompressing compressed data obtained by a data compression device in which all of the blend contexts are parallelized, the code separation / selection unit in FIG. 60 is the embodiment in FIG. 64, and the word stream decoding unit 702 in FIG. The byte stream decoding unit 704 is the embodiment of FIG. 65, and the output switching unit 706 is the embodiment of FIG. Further, this embodiment exemplifies the restoration of compressed data in the case of a one-word 4-byte configuration, and further exemplifies a case in which a selection code is added in byte units as shown in FIG.

The code separation / selection section 700 shown in FIG. 64 has a code separation section 708 at the input stage, receives the compressed data, separates the number of bytes constituting one word, that is, the code of 4 bytes, and Code buffer units 710-1 to 710-7 provided for the first, second, third, and fourth bytes, respectively.
10-4. This separated code is shown in FIG.
As shown in (A), a selection code is provided at the beginning, and code data to be compressed byte data is provided after the selection code.

Following the code buffer units 710-1 to 710-4, the selective code decoding units 712-1 to 712-4,
Word / byte selection probability model units 716-1 to 716-
4 and code switching units 714-1 to 714-4 are provided in parallel for each byte position. For example, taking the first byte as an example, the selection code decoding unit 712 inputs the first selection code of the code data of the first byte stored in the code buffer unit 710-1, and outputs the selected code immediately before the restoration. The conditional ease is determined from the code string, and whether the selected code indicates a word stream code or a byte stream code is decoded.

Word / byte selection probability model unit 716-
Numeral 1 stores frequency data in which the conditional easiness of the input selection code connected to the immediately preceding selection code string is registered. The frequency data is searched by the conditional ease calculated by the selection code decoding unit 712-1. To restore the selected code, and further updates the frequency data according to the result of the restoration. The code switching unit 714-1 converts the code stored in the code buffer unit 710-1 into a word stream code W1 or a byte stream code B1 according to the decoding result of the selection code decoding unit 712-1 in FIG.
Of the word stream decoding unit 702 and the first byte of the byte stream decoding unit 704. The remaining second byte, third byte, and fourth byte in FIG.
The same applies to the byte.

A word stream decoding unit 702 shown in FIG.
The byte stream decoding unit 704 decodes the word stream code or the byte stream code at each byte position separated and selected by the code separation / selection unit 700 in FIG. 64 in parallel. The word stream decoding unit 702 corresponds to each of the word stream codes W1 to W4 of the first byte, the second byte, the third byte, and the fourth byte.
20-1 to 720-4, decoding units 722-1 to 722-
4 and the probability model units 724-1 to 724-4.

The decoding units 722-1 to 722-4 among them
, Since decoding from the 0th order to the 3rd order is handled as the blend context, 0th to 3rd order encoding units are provided in parallel. Similarly, the probability model units 724-1 to 724
-4 also includes the 0th to 3rd order probability model units in parallel. Code distribution units 720-1 to 720-
4 is each input code W1 to W4 of the 1st to 4th bytes
From the bit configuration of 0, 1 order, 2 order,
Alternatively, a third order is recognized, and the corresponding decoding units 722-1 to 72-7 are recognized.
22-4. Decoding sections 722-1 to 722 for decoding all orders of the blend context in parallel
-4 and the details of the probability model units 724-1 to 724-4 are as shown in FIGS.

On the other hand, like the word stream decoding unit 702, the byte stream decoding unit 704 adopts an embodiment in which decoding of all orders is performed in parallel for the blend context. That is, the code distribution unit 730- corresponds to the input byte stream codes B1 to B4 of the first to fourth bytes.
Decoding units 732-1 to 732-4 having decoding units of 0th to 3rd order in parallel and 0th to 3rd order
734-1 to 734 provided with the following probability models in parallel
-4.

[0284] Such a word stream decoding unit 70
64 and the byte stream decoding unit 704 in FIG.
The four codes of 4 bytes constituting one word selected by the code separation / selection unit 700 at the preceding stage provided as input signals are selectively input to the word stream decoding unit 702 and / or the byte stream decoding unit 704 as effective input codes. The byte data for one word is restored in parallel.

In the word unit restoration, four of the input codes W1 to W4 and B1 to B4 to the word stream decoding unit 702 and the byte stream decoding unit 704 are valid input codes, and , And decoding is prohibited for a system without a valid input code. The output switching unit 706 in FIG. 66 includes the byte data decoded in parallel by the word stream decoding unit 702 and the byte stream decoding unit 704 in FIG. 65, that is, the restored byte data WR11 to WR41 in the word stream direction and the byte stream direction. Of the restored byte data BR11-BR41 for each byte position from the first byte to the fourth byte, and outputs the restored byte data on the side decoded with valid code input to the output selector 84.
0-1 to 840-4 are selected and input to the context capture register unit 842.

The context capturing register section 842 includes register groups 842-1 to 842-4 in which four registers are connected in series for each byte position. Here, the register group 842-1 of the first byte includes the registers R110 to R1.
The register group 842-2 at the second byte is indicated by registers R210 to R240, the register group 842-3 at the third byte is indicated by registers R310 to R340, and the register group 842- at the fourth byte. 4 is indicated by registers R410 to R440.

The register group R110, R210, R310, R410 of the first stage of the register unit 842 for context capture
Holds the restored byte data of each byte position for one word. Therefore, the registers R110 to R110
By reading each byte data of R410 in parallel,
One-word word data is obtained as restoration source data.

At this time, the second-stage registers R120 to R4
20, the third-stage registers R130 to R430 and the fourth-stage registers R140 to R440 store the byte data of each word restored one before, two before, and three before each byte position. Is held. The byte data of each of these already restored words is supplied to the word stream decoding unit 702 and the byte stream decoding unit 704 in FIG.
Is included as a context in the probability model part of.

With respect to the capture of the context, the probability model unit 724 provided in the word stream decoding unit 702
Regarding −1 to 724-4, byte data in the word stream direction is captured for the byte data stored in the context capturing register unit 842 in FIG. 66 for the first to third-order probability model units. . For example, taking the probability model unit 742-1 at the first byte as an example, FIG.
Since the restored byte data of the 1st byte before one word, two words before, and three words before are stored in the registers R120, R130, R140 of FIG. , R140
As a context in the first to third order probability model parts.

On the other hand, with respect to the probability model units 734-1 to 734-4 of the byte stream decoding unit 704 in FIG. 65, byte data in the byte stream direction is captured in the context capturing register unit 842. I have. For example, taking the probability model unit 734-4 at the fourth byte as an example, registers R310, R210, and R110 are used as byte data in the byte stream direction preceding the current restored byte data for the zero-order probability model unit.
Is taken as byte data for the primary, secondary, and tertiary contexts.

Of course, the word stream decoding unit 70
2 probability model units 724-1 to 724-4 and the probability model units 734-1 to 734-1 of the byte stream decoding unit 704.
In any of the cases 34-4, since the restoration of the word data into the 0th-order probability model unit uses the restored data by the 0th-order decoding unit at the preceding stage as it is, the byte data from the context acquisition register unit 842 in FIG. No uptake required.

Although the encoding and decoding in the word stream direction in the above-described embodiment is an example in which word data is divided into a plurality of byte data and compressed or decompressed, the present invention is not limited to this. Instead, it is also possible to separate the operation code for each type of a plurality of bits in the operation code of a program or the like, and to perform encoding and decoding for each separated data unit.

Further, in the above-described embodiment, as an example, the coding and decoding of the probability statistical type is taken as an example. In addition to this, the dictionary-type decoding and coding represented by LZW and LZSS are completely the same. Can be applied.

[0294]

As described above, according to the present invention, the probability model unit and the entropy encoding unit are operated in a pipeline for data compression in a fixed context, and the entropy decoding unit and the probability model are used for data restoration in a fixed context. By performing the pipeline operation of the units in the same manner, it is possible to realize the coding and decoding of the probability statistical type of fixed-order context, which has conventionally taken a long time by complicated processing, at high speed.

Also, regarding the compression and decompression of the blend context,
By parallelizing and pipelining at least the stochastic model part, it is possible to compress and decompress the probability statistical code for the blend context, which took time due to the serial processing for each order, in a practically usable processing time and at a higher speed. can do. Further, according to the present invention, by performing encoding using a dictionary type or a probability statistical type in byte units in the word stream direction, 2-byte Unicode or Japanese code,
In encoding for compressing word data such as a program code such as a full-color image of 3 bytes of RGB, 4 bytes, and 8 bytes, the compression efficiency can be increased by incorporating the correlation of the same byte position in the word data into the encoding. Also in this case, compression and decompression in which the use time is reduced to a practical level can be realized by pipelining the decoding and encoding processes.

Further, as a decompression of the word data, decoding is performed in a word stream direction and a byte stream direction for each byte data constituting the word data, and a code having a higher compression efficiency is adaptively selected. Efficient compression processing that takes advantage of both the encoding in the word stream direction and the encoding in the byte stream direction can be realized.

[Brief description of the drawings]

FIG. 1 is a diagram illustrating the principle of the present invention.

FIG. 2 is a diagram illustrating the principle of the present invention (continued)

FIG. 3 is a diagram illustrating the principle of the present invention.

FIG. 4 is a diagram illustrating the principle of the present invention (continued).

FIG. 5 is a block diagram of a basic configuration of a data compression device pipelined for a fixed-order context.

FIG. 6 is a block diagram of a data compression apparatus employing arithmetic coding as the entropy coding of FIG. 5;

FIG. 7 is a block diagram of a detailed configuration of the data compression device of FIG. 6;

FIG. 8 is an explanatory diagram of a frequency data storage unit in FIG. 7;

9 is an explanatory diagram of an output code format of the arithmetic coding of FIG. 7;

FIG. 10 is an explanatory diagram of a comparison process of a section [1, 0] for generating an output code in FIG. 9;

FIG. 11 is a time chart of the pipeline operation of FIG. 7;

FIG. 12 is a block diagram of a basic configuration of a data restoration apparatus pipelined for a fixed-order context.

FIG. 13 is a block diagram of a data restoration apparatus employing arithmetic decoding as the entropy decoding of FIG. 12;

FIG. 14 is a block diagram of a detailed configuration of the data restoration device in FIG. 13;

FIG. 15 is a time chart of the pipeline operation of FIG. 14;

FIG. 16 is a time chart of the pipeline operation of FIG. 14 (continued)

FIG. 17 is an explanatory diagram of a basic block and a processing operation of a data compression device in which a blend context is partially parallelized;

18 is a block diagram of a detailed configuration of the data compression device in FIG.

FIG. 19 is a time chart of the pipeline operation of FIG. 18;

FIG. 20 is an explanatory diagram of basic blocks and a processing operation of a data restoration device in which a part of a blend context is partially parallelized;

FIG. 21 is a block diagram of a detailed configuration of the data restoration device in FIG. 20;

FIG. 22 is a time chart of the pipeline operation of FIG. 21;

23 is a time chart of the pipeline operation in FIG. 21 (continued)

FIG. 24 is an explanatory diagram of basic blocks and processing operations of a data compression device in which all blend contexts are parallelized;

FIG. 25 is a block diagram of a detailed configuration of the data compression device in FIG. 24;

FIG. 26 is a time chart of the pipeline operation of FIG. 25;

FIG. 27 is an explanatory diagram of basic blocks and a processing operation of a data restoration device in which all blend contexts are parallelized;

FIG. 28 is a block diagram of a detailed configuration of the data restoration device in FIG. 27;

FIG. 29 is a time chart of the pipeline operation of FIG. 28;

FIG. 30 is a block diagram of an internal hash structure of a frequency data storage unit provided in an appearance frequency model unit according to the present invention.

FIG. 31 is a block diagram of an internal hash structure and an external hash structure of a frequency data storage unit provided in an appearance frequency model unit according to the present invention.

FIG. 32 is a block diagram of a data compression device that performs encoding in byte units in the word stream direction.

FIG. 33 is a time chart of the encoding process in FIG. 32;

FIG. 34 is a flowchart of an encoding process for compiling compressed byte data in word units.

FIG. 35 is an explanatory diagram of a format of compressed data in which compressed byte data is collected in word units.

FIG. 36 is an explanatory diagram of encoding in the word stream direction taking a 1-word 4-byte configuration as an example.

FIG. 37 illustrates the compressed byte data of FIG. 36;
Illustration of format 5

FIG. 38 is a flowchart of an encoding process for combining compressed byte data in file units;

FIG. 39 is an explanatory diagram of a format of compressed data in which compressed byte data is collected in file units;

FIG. 40 is a diagram showing the compressed byte data of FIG. 36 as a target;
Specific explanatory diagram of format 9

FIG. 41 is a block diagram in the case where probability statistical coding is applied to FIG. 32;

FIG. 42 is a time chart of the encoding pipeline processing of FIG. 41;

FIG. 43 is a block diagram of a data restoration apparatus that decodes data in byte units in the word stream direction.

FIG. 44 is a time chart of the decoding process in FIG. 43;

FIG. 45 is a flowchart of a process of decoding compressed data in which compressed byte data is grouped in word units;

FIG. 46 is a flowchart of compressed data decoding processing in which compressed byte data is collected in file units;

FIG. 47 is a block diagram in the case where probability statistical decoding is applied to FIG. 43;

FIG. 48 is a time chart of the decoding pipeline processing of FIG. 47;

FIG. 49 is an explanatory diagram of pipeline control when arithmetic coding is applied to byte data in the word stream direction.

FIG. 50 is an explanatory diagram of pipeline control when arithmetic decoding is applied to byte data in the word stream direction.

FIG. 51 is an explanatory diagram of pipeline control when code tree coding is applied to byte data in the word stream direction.

FIG. 52 is an explanatory diagram of pipeline control when arithmetic decoding is applied to byte data in the word stream direction.

FIG. 53 is a block diagram of a data compression device that adaptively switches encoding in a word stream direction and a byte stream direction.

FIG. 54 is a time chart of the encoding process in FIG. 53;

FIG. 55 is an explanatory diagram of a format of compressed data obtained in FIG. 54;

FIG. 56 is a flowchart of an encoding process when a switching selection code is added in units of compressed byte data in FIG. 54;

FIG. 57 is a flowchart of an encoding process when a switching selection code is added in units of compressed words in FIG. 54;

FIG. 58 is a block diagram of an embodiment of a byte array conversion unit, a word stream coding unit, and a byte stream coding unit in FIG. 53 in the case of applying the probability statistical coding that makes the blend context all parallel;

FIG. 59 is a block diagram of an embodiment of the switching output unit in FIG. 53 in the case of applying the probability statistical coding for fully parallelizing the blend context.

FIG. 60 is a block diagram of a data restoration device that adaptively switches decoding in the word stream direction and the byte stream direction.

FIG. 61 is a time chart of the decoding process in FIG. 60;

FIG. 62 is a flowchart of a decoding process of compressed data in which a switching selection code is added in units of compressed byte data in FIG. 60;

FIG. 63 is a flowchart of a process of decoding compressed data to which a switching selection code is added in units of compressed words in FIG. 60;

FIG. 64 is a block diagram of an embodiment of the code separation / selection unit in FIG. 60 in the case of applying the probability statistical decoding that makes the blend contexts all parallel.

FIG. 65 is a block diagram of an embodiment of the word stream encoding unit and the byte stream encoding unit in FIG. 60 in the case of applying the probability statistical coding for fully parallelizing the blend context.

FIG. 66 is a block diagram of an embodiment of a switching output unit in FIG. 60 in the case of applying probability statistic type decoding that makes the blend contexts all parallel;

FIG. 67 is a block diagram of a basic configuration of a conventional probability statistical data compression apparatus.

FIG. 68 is an explanatory diagram of a tree structure of a context and a conditional probability with respect to a character string;

FIG. 69 is a view for explaining the principle of conventional multi-level arithmetic coding.

FIG. 70 is a functional block diagram of a conventional multi-level arithmetic coding.

FIG. 71 is a flowchart of conventional multi-level arithmetic coding using a single history (zero-order fixed context).

FIG. 72 is a flowchart of conventional multi-level arithmetic coding using double history (primary fixed context).

[Explanation of symbols]

10: Appearance frequency model part 11, 15, 47, 51, 85, 125, 175, 21
5: Pipeline control unit 12: Entropy coding unit 14: Probability model unit 16: Arithmetic coding unit 18, 94: Rank reading / updating unit 20, 72, 98, 156: Character reading / updating unit 22, 74, 100, 158 : Frequency read / update units 24, 76, 100, 160: cumulative frequency read / update units 25, 70, 96, 154: frequency data storage units 26, 78, 106, 162: total cumulative frequency read / update units 28, 58, 116, 142: code space calculation unit (high level calculation unit) 30, 60, 118, 144: code space calculation unit (low level calculation unit) 32, 62, 66, 120, 146, 150: reciprocal calculation unit 34, 122: arithmetic Code deriving units 36, 124, 138, 170: Code output units 38-1 to 38-4, 40-1 to 40-3, 42-1, 42-2, 44: Code space of the last character 46: Entropy decoding Part 48: Frequency of appearance Model part 50, 126, 224: arithmetic decoding part 52: probability model part 54, 212: code input part 56, 140: cumulative value deriving part 64, 148: cumulative frequency deriving part 68, 152: rank deriving part 80, 164 : Character output unit 82, 128, 170, 216, 166: Blending probability model unit 84: Buffer selection unit 86, 168: Arithmetic coding unit 88, 130, 172, 230: 0th-order probability model unit 90, 132, 174 , 232: first-order probability model unit 92, 134, 176, 234: second-order probability model unit 95: escape frequency search / update unit 104: escape determination unit 108, 110, 112: buffer unit 114, 136, 136-1 136-3,190,2
36: Selection unit 178, 172: 0th-order arithmetic coding unit 180, 174: 1st-order arithmetic coding unit 182, 176: Secondary-arithmetic coding unit 184, 218: 0th-order code buffer unit 186, 220: 1st-order code Buffer unit 188, 222: Secondary code buffer unit 190: Merging unit 192 :, 194, 196: Rank reading / updating unit 198, 200, 202: Frequency / cumulative frequency reading / updating unit 204, 258: Selection unit 211: Separation unit 260 , 262, 264: code space calculation unit 266, 268, 270: arithmetic code derivation unit 272, 274, 276: serial / parallel conversion code output unit 211: separation unit 224, 172: zero-order arithmetic decoding unit 226, 174: Primary arithmetic decoding unit 228, 176: Secondary arithmetic decoding unit 278, 280, 282: Parallel / serial conversion code input unit 28 , 286, 288: Arithmetic code rank deriving unit 290, 292, 294: Code space calculation unit 246, 248, 250: Rank / frequency / cumulative frequency reading / updating unit 246, 248, 250: Character reading / updating unit 500: Byte array conversion Units 502, 502-1 to 502-n: Byte unit encoding unit 503: Code merging unit 504, 514: Header 504-1 to 504-n: Byte header 506, 516: Compressed data 512-1, 512-2: Compressed word data 514-1 to 514-n: file header 516-1 to 516-n: byte file 518-1 to 518-4, 540-1 to 540-4: register group (for capturing context) 520-1 520-4: Appearance frequency model section 522-1 to 522-4: Entropy coding section 524-1 to 524-4, 534-1 to 34-4: Code buffer unit 530: Code separation unit 530, 532-1 to 532-n: Byte unit decoding unit 530: Code separation unit 536-1 to 536-4: Entropy decoding unit 538-1 to 538- 4: appearance frequency model unit (for restoration) 542: probability model search unit 544: probability model update unit 546: code space operation unit 548: arithmetic code output unit 550: arithmetic code input unit 552: cumulative value operation unit 554: code space Arithmetic unit 556: Probability model search unit 558: Output character search unit 560: Probability model update unit 562: Context collection processing unit 564: Code tree coding unit 568: Context tree search unit 570: Context tree registration unit 572: Same context judgment Unit 574: code tree search unit 576: code tree update registration unit 578: code rearrangement unit 580: code tree decoding unit 582: context collection processing unit 58 : Code tree search unit 588: code tree update registration unit 590: context tree search unit 592: context tree registration unit 600: byte array conversion unit 602: word stream coding unit 604: byte stream coding unit 606: code switching unit 620 -1 to 620-4, 630-1 to 630-4: Probability model unit 622-1 to 622-4, 632-1 to 632-4: Encoding unit 624-1 to 624-4, 634-1 to 634 -4: code buffer section 640-1 to 640-4: bit number comparing section 642-1 to 642-4: word / byte selection probability model section 644-1 to 644-4: coding section 648-1 to 648- 4: code switching unit 652: code merging unit 700: code separation selection unit 702: word stream decoding unit 704: byte stream decoding unit 706: output switching unit 708: Code separation units 712-1 to 712-4: selected code decoding units 714-1 to 714-4: code switching units 716-1 to 716-4: word / byte selection probability model units 720-1 to 720-4, 730-1 to 730-4: code distribution units 722-1 to 722-4, 732-1 to 732-4: decoding units 724-1 to 724-4, 734-1 to 734-4: probability model units 840 -1 to 840-4: output selection unit

Continuation of front page (56) References JP-A-2-202267 (JP, A) JP-A-4-280517 (JP, A) JP-A-5-128103 (JP, A) JP-A-5-14206 (JP) JP-A-5-224878 (JP, A) JP-A-5-128100 (JP, A) JP-A-5-241777 (JP, A) JP-A-6-28149 (JP, A) 6-90362 (JP, A) JP-A-6-44038 (JP, A) JP-A-6-149537 (JP, A) JP-A-3-213019 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) H03M 7/40

Claims (11)

(57) [Claims] In a data compression apparatus for encoding an input character according to a conditional ease of a preceding character string, frequency data in which the conditional ease of an input character connected to the immediately preceding character string is registered for each character An appearance frequency model unit that has a storage unit, searches for the conditional ease of input characters connected to the immediately preceding character string by referring to the frequency data storage unit, and registers and updates the frequency data storage unit; comprising an entropy encoding unit encoding an input character by the conditional Deeki of determined model unit, a pipeline control unit for the occurrence frequency modeling unit and entropy coding unit pipelined operation, said pipeline In the first cycle, the control unit
Input the first character of the column to the appearance frequency model
The conditional ease of character output is determined, and each
In the cycle, the next character of the character string is
Input to the text box to determine the conditional ease of input characters.
In addition, the conditional ease of character obtained in the previous cycle
Appearance frequency of the input character input to the entropy encoding unit
A data compression apparatus that repeats the encoding process . 2. A data restoration apparatus for restoring a character according to a conditional ease of input of an inputted code and a character string immediately before being restored. An entropy decoding unit that obtains a conditional ease from the input data, and a frequency data storage unit that registers, for each character, the conditional ease of an input character connected to the immediately preceding character string, and a condition obtained by the entropy decoding unit. The frequency data storage unit is searched according to the ease of attachment to restore characters, and an appearance frequency model unit for registering and updating the frequency data storage unit is pipelined with the entropy decoding unit and the appearance frequency model unit. And a pipeline control unit, wherein the pipeline control unit receives an input character in a first cycle.
The first character of the column is input to the entropy decoding unit and
The conditional ease of input characters is determined, and the second and subsequent cycles
In each cycle, the next character of the string is
To the conditional decoding of the input character
As well as conditional access to characters determined in the previous cycle
Is input to the appearance frequency model section to restore the input characters.
A data restoring apparatus characterized by repeating a process of performing a data restoration. 3. A data compression apparatus for encoding an input character from a high-order context consisting of a long character string immediately before to a low-order context consisting of a short character string according to conditional easiness of output. A frequency data storage unit in which the conditional ease of input characters connected is registered for each of the orders, and the conditional ease of input characters connected to the immediately preceding character string in each order is referred to by referring to the frequency data storage unit. A plurality of appearance frequency model units provided for each order, in which a search and a process of registering and updating the frequency data storage unit are performed in parallel for each order, and a condition for each order of the plurality of appearance frequency model units An entropy encoding unit that encodes the ease of attachment from a higher-order context to a lower-order context, the plurality of appearance frequency model units, and the entropy encoding
Includes a pipeline control unit for the parts pipeline operation, said pipeline control unit, an input character in a first cycle
The first character of the column is added in parallel to the plurality of appearance frequency model parts.
To obtain the conditional ease of output of each order in parallel.
After the cycle, the conditions for each order obtained in the first cycle
The degree of attachment is output to the entropy encoding unit, and
Arithmetic encoding is performed in the order of pulse code and low-order context code.
Sometimes, the next character of the input character string is
Input to the model part and calculate the appearance frequency of each number of characters in parallel
A data compression device characterized by repeating processing . 4. A data restoration apparatus for restoring characters in order from a high-order context consisting of a long character string immediately before being restored with an inputted code and a short character string. An entropy decoding unit that obtains conditional ease in order from higher-order context to lower-order context based on the code and the character string immediately before each degree that has already been restored; and conditional ease of input characters connected to the immediately preceding character string. Has a frequency data storage unit registered for each of the degrees, and restores a character by referring to the frequency data storage unit from the conditional ease of each degree obtained by the entropy decoding unit, and stores the frequency data storage unit A plurality of appearance frequency model units provided for each order to perform a process of registering and updating in parallel for each order, and the entropy decoding unit and a plurality of appearance frequency model units.
The includes a pipeline control unit for pipeline operation, and the pipeline control unit, an input code in the first cycle
The head code of the column is input to the entropy decoding unit.
The conditional appearance easiness of the order is calculated to obtain the plurality of appearance frequency models.
Output to the controller in parallel.
Corresponds to the degree of conditional ease output in the first cycle
Output the characters restored by the appearance frequency model part
At the same time as the next code of the input code sequence
Determining the conditional ease of order that has been input to the decoding unit,
Repeat the process of outputting to multiple appearance frequency model parts in parallel.
A data restoration device characterized by returning . 5. A data compression apparatus for encoding an input character from a high-order context consisting of a long character string immediately before to a low-order context consisting of a short character string by conditional easiness of output. A frequency data storage unit in which the conditional ease of input characters connected is registered for each of the orders, and the conditional ease of input characters connected to the immediately preceding character string in each order is referred to by referring to the frequency data storage unit. A plurality of appearance frequency model units independently provided for each order, in which a search and a process of registering and updating the frequency data storage unit are performed in parallel for each order, and each order of the plurality of appearance frequency model units. The entropy encoding of the conditional ease of execution is performed in parallel for each order.An entropy encoding unit provided independently for each order, and a code output from the entropy encoding unit for each order. collect A code merging unit for force, said plurality of frequency model provided in parallel for each order
And a plurality of entropy coding units for each order
Pipeline and a pipeline control unit for operating the pipeline control unit, out of the first character of the input string of the plurality in the first cycle to
Conditional output of each order by parallel input to the current frequency model
Are calculated in parallel, and the plurality of occurrence frequencies corresponding to the respective orders are calculated.
Output to each of the degree model parts , and for the second and subsequent cycles, each order obtained in the first cycle
The plurality of entropy codes corresponding to conditional ease of
Input to encoding unit to encode from higher-order context to each code in lower-order context
At the same time, the next character of the input character string is
Input to multiple appearance frequency model sections to calculate the appearance frequency of each character
The plurality of appearance frequency models corresponding to each order are obtained in parallel.
A data compression device for repeating a process of outputting to each of the Dell units. 6. A data compression apparatus according to claim 5, wherein:
The pipeline control unit includes the plurality of appearance frequency models.
If the conditional accessibility of each order is required for each of the
Conditional output determined by the appearance frequency model part of the higher-order order
Easiness output, and other low-order occurrence frequency models
To prohibit the output of conditional ease that was requested by Dell
Characteristic data compression device. 7. A data restoration apparatus for restoring characters in order from a high-order context consisting of a long character string immediately before being restored with an inputted code and a short character string, which has already been restored. A code separation unit that separates a code into codes of each order, and a process of obtaining conditional ease of output of each order based on the code of each order separated by the separation unit and a character string immediately before each order that has been restored. , An entropy decoding unit provided independently for each order that performs each order in parallel, and a frequency data storage unit that registers the conditional ease of input characters connected to the immediately preceding character string for each of the orders A process of retrieving the frequency data storage unit and restoring characters by retrieving the character with the conditional ease of each order obtained by the plurality of entropy decoding units and registering and updating the frequency data storage unit; Provided for each order performed for each A plurality of occurrence frequency modeling unit that, said plurality of entropy provided in parallel for each order
The decoding unit and the plurality of appearance frequency model units are
A pipeline control unit for performing a pipeline operation , wherein the pipeline control unit is configured to perform the pipeline operation in a first cycle.
Each of the constituents of the leading code of the input code string separated by the separation unit
A code of order is sent to the plurality of entropy decoding units.
Entering in a row and finding conditional ease of said multiple occurrence frequencies
Output to the model part in parallel.
The conditional ease of each order determined in the first cycle
Input to the plurality of appearance frequency model parts corresponding to
At the same time as outputting the characters restored by the frequency model part,
Constructs the next code of the input code string separated by the code separation unit
The plurality of codes of each order
Section in parallel to obtain the conditional output
A data restoration apparatus characterized by repeating a process of outputting to a current frequency model part in parallel . 8. The data restoration device according to claim 7, wherein:
The pipeline control unit, separated by the sign-separating unit
One or more of the above-mentioned one or more
Allow the output of conditional ease determined by the entropy decoding unit.
And other outputs of the entropy decoding unit are prohibited.
A data restoration device characterized by stopping. 9. In the data compression apparatus and a data restoration apparatus according to any one of claims 1 to 8, wherein the occurrence frequency modeling unit as information for directing Deeki of the conditional,
Search for rank, number of occurrences, cumulative number of occurrences, total number of occurrences in context,
A data compression device and a data decompression device which perform registration and updating, and wherein the entropy coding unit and the entropy decoding unit perform arithmetic coding and arithmetic decoding. 10. The data compression device and the data decompression device according to claim 9 , wherein the pipeline control unit determines the order of appearance, the number of occurrences, and the cumulative occurrence of the conditional appearance in the appearance frequency model unit. A data compression device and a data decompression device characterized in that reading, addition and writing of the total number of occurrences under a number and a context are pipelined for each information. 11. A data compression apparatus according to claim 10 , wherein said pipeline control unit performs an operation of an upper end and an lower end representing an arithmetic code space in parallel. And a data restoration device.