JPH04280517A - Data compression and restoring system - Google Patents

Abstract

PURPOSE:To attain efficient data compression and restoration by an arithmetic coding system to multi-value data in a data compression system applying arithmetic coding to a data such as a character code or a picture data. CONSTITUTION:The data compression and restoring system in which the inputted character string is coded by an arithmetic code and data compression and restoration are implemented is provided with an SOR coding section (1) rearranges a character string with a higher frequency of occurrence to have a smaller registration number in a dictionary every time a character string is entered and coding the character string according to the registration number in the character string and an arithmetic coding section (2) receiving a code outputted from the SOR coding section (1) and applying the arithmetic coding to the character string and is featured to encode the input character string with the SOR code and encode the code with the arithmetic code.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a data compression method for arithmetic encoding data such as character codes and image data. As computers begin to handle various types of data such as character codes and image data, the amount of data handled is also increasing. When processing such large amounts of data, in order to reduce storage capacity or enable transmission to remote locations, it is necessary to omit redundant parts of the data and compress it for storage or data transfer. It is desirable to do so.

[0002] As a data compression method that can be applied to various types of data (character codes, image data, etc.), there is a method using universal codes. There are various universal encoding methods, but a typical one is the arithmetic encoding method. An object of the present invention is to provide a method for efficiently compressing and restoring multivalued data using an arithmetic coding method.

0003

BACKGROUND OF THE INVENTION Arithmetic encoding systems can achieve maximum compression when the frequency of appearance of characters in an information source is known. Universal codes can be applied not only to character codes in document data, but also to compression of data such as image data.In the following explanation, one word of data is referred to as a character, and a string of arbitrary words of data is referred to as a character. It's called a string.

Multivalued arithmetic codes will be explained with reference to FIG. FIG. 7(A) shows the appearance frequency of each symbol (character) when the input characters consist of four characters a, b, c, and d. Figure (B) shows the symbols and the cumulative appearance frequency. Figure (C) shows arithmetic encoding. Multi-value arithmetic is an encoding method that associates a character string of input data with one point on a number line [0, 1] and represents it with one code. As an example of encoding, for the sake of simplicity, consider an information source consisting of only four characters, a, b, c, and d.

Assume that the appearance frequency of each symbol (probability is calculated by dividing the total number of appearances by the number of appearances for each symbol) is as shown in FIG. Next, the symbols are rearranged in order of frequency, and the cumulative probabilities are calculated in descending order of frequency, as shown in Figure (B).
c, which has the highest frequency, accumulates the appearance frequency (probability) of b, d, and a, b accumulates the appearance frequency of d and a, and d accumulates the appearance frequency of a).

[0006] Therefore, when symbol (character) c is input, as shown in Figure (C), an interval 50 obtained by dividing the interval [0, 1] according to the cumulative frequency of c is selected. Next, when symbol a is input, interval 50 is the cumulative appearance frequency of each symbol a (the maximum value of cumulative appearance frequency is normalized to 1)
The section 51 corresponding to the symbol a obtained by is selected. Next, when symbol d is input, an interval 52 divided by the cumulative frequency of symbol d is selected from interval 51, and an arbitrary value of interval 52 (any value between the upper and lower ends of interval 52) is selected.
Outputs as a sign.

The lower end of the interval is determined by the following equation. Lower end of new subinterval = Lower end of current subinterval + Current subinterval width × Cumulative probability of appearing characters New subinterval width = Current subinterval width × Probability of appearing characters Also, in codeword restoration, each codeword is The original string of characters (symbols) can be restored by sequentially re-dividing and checking which interval the characters are included in based on their probabilities.

According to multilevel arithmetic coding, the more frequently a character string appears, the larger the width of the final section becomes, and it can be represented by a shorter code, thereby compressing the amount of data. Furthermore, in this method, there is no restriction on the minimum unit for the bit representation of the code word, and the minimum unit can be set to one bit or less, so highly efficient compression can be achieved.

FIG. 8 shows the configuration of an arithmetic encoding device. In the figure, 60 is a frequency sorting unit, 61 is a counter that counts frequencies, 62 is a dictionary consisting of symbols and identification numbers sorted in frequency order, and 63 is a cumulative appearance frequency output from the counter and a frequency sorting unit. This is an arithmetic code section that performs arithmetic coding using the identification number output from the .

The operation of the configuration shown in the figure will be explained. Each time one character is input from the input character string, the counter 61 counts the frequency of appearance. The frequency sorting unit 60 sorts the symbols in order of appearance frequency and rewrites the dictionary 62. and,
The arithmetic code unit 63 calculates the upper and lower ends of the interval based on the cumulative frequency of the input characters, the interval width and the cumulative appearance probability, and calculates the upper and lower ends of the interval.
Outputs any value in the interval as a sign.

FIGS. 9 and 10 show the flow of arithmetic encoding. Initial values are upper end = 1, lower end = 0, section width = 1.0
shall be. cum[i] represents the appearance frequency of identification number i (the i-th appearance frequency sorted in order of magnitude). cum
freq[i] is the i-th cumulative frequency, cum
freq[i] represents the cumulative appearance frequency from cum[i] to cum[N]. cum freq[1]=1
, cum freq[N]=0.

FIG. 9 shows the flow of multilevel arithmetic encoding when the history of appearing characters is not considered (single history). S1 (1) Assign all single characters to each slot of dictionary D. D(i)=i(i=1,2,...A),
However, A is the size of the alphabet (the number of alphabets and symbols, etc.), which can be up to 256. (2) Assign a number to each character. I(i)=i(
i=1, 2, ..., A). (3) Initialize the appearance frequency. freq(i)=1
(i=1, 2,...,A). (4) Initialize the cumulative appearance frequency. cum fr
eq(i) = A-1 (i = 1, 2, ..., A, the cumulative frequency of appearance is initialized like this because all values must be different.) S2 Input one character k do. S3 Find the slot (section) of the input character using j=I(k) and i=D(j), and arithmetic encode the number j.

S4: Replace the characters corresponding to the slots in order of appearance frequency. Among the symbols whose number is smaller than j, find a symbol whose appearance frequency is different from the j-th symbol. That is, freq(r)! Find the largest r(<j) such that = freq(j) (! means different). Furthermore, suppose that the identification number of the character in the r-th slot of appearance frequency is I(r)=s, and the identification number of the character corresponding to the r-th slot of cumulative appearance frequency is t ((D(
r))=t). Therefore, the r-th and j-th are exchanged. That is, I(j)=S (character of identification number s), D(j)=
t, I(r)=j, D(r)=i(=D(j)). S5 Increment the appearance frequency and cumulative appearance frequency by 1. However, the cumulative appearance frequency is incremented by 1 for all slots below r. Repeat S2 and subsequent steps.

FIG. 10 shows a flow when double history is taken into account for characters that appear. When considering double history, consider the combination of two consecutive characters, create a dictionary for each character, and record the frequency of occurrence of the character following each character for each second character. Make it. For example, in a dictionary for the character a, the appearance frequency of each character following a (such as b for ab, c for ac, etc.) is determined and recorded.

Therefore, in the slot of dictionary D, for each character i following character p, D(p, i) (p, i=1, 2
,...,A) (A is the size of the alphabet (number))
Assign. Also, the number assigned to each character is I(p
, i) (p, i=1, 2, . . . , A). Other than the above points, the flow is no different from the flow in FIG. 9, so the explanation will be omitted.

[0016]

[Problems to be Solved by the Invention] As mentioned above, in arithmetic encoding, it is necessary to rearrange the frequency of appearance each time a character is input, and this process takes a long time. Therefore, although conventional arithmetic coding can achieve a high compression ratio, it is inefficient in terms of time. An object of the present invention is to provide a data compression and restoration method that performs arithmetic encoding at high speed.

[0017]

[Means for Solving the Problems] The present invention first uses an encoding method (hereinafter referred to as SOR code) based on self-organizing rules in which frequently appearing characters are always encoded with a small number to encode input characters. The sequence is encoded, and the SOR code is arithmetic encoded.

[0018] Prior to explaining the present invention, the SOR
The symbols will be explained. In the SOR encoding method, each character string that appears in input data is registered in a dictionary, and repeated character strings are represented and encoded by the registration number of the same character string in the dictionary, and the registration number is assigned to the character strings that appear more frequently. The dictionary is dynamically updated each time a character is input so that the number becomes smaller.

FIG. 11 shows an encoding method (SOR encoding method) based on self-organizing encoding rules. Figure (A) shows the MTF (Move To Front) self-assembly method.
) show the law. There are four types of input characters, a, b, c, and d, and a case will be considered in which the illustrated input data 'abacbc...ad' is encoded. (1) Output the first character a as code 1a and register it as number 1 in the dictionary. (2) Output the second character b with the code 2b, lower the character a with the registration number 1 to the registration number 2, and register b with the registration number 1 in the dictionary. (3) Output the third character a with code 2. and,
In the dictionary, a is registered with registration number 1, and b is registered with registration number 2.

(4) Output the fourth input character c with the symbol 3c. Then, c is registered with registration number 1, a with registration number 2, and b with registration number 3 in the dictionary. (5) Output the fifth character b as 3. Then, in the dictionary, b is registered with registration number 1, c with registration number 2, and a with registration number 3. Similarly, the characters that appear are encoded using the registration number of the dictionary, the dictionary is updated, and the characters that appear are registered with the registration number 1.

FIG. 11(b) shows the TR (Transpose) method among the self-assembly methods. The input character is a,
There are four types, b, c, and d, and the character string 'abac' as shown in the diagram
The case where b...ad' is encoded will be described. (1) Output the first character a with the code 1a, and register a with the registration number 1 in the dictionary. (2) Output the second character b with code 2b, and register b in the dictionary with registration number 2. (3) Output the third character a with the code 1. (4) Output the fourth character c with code 3c and register c in the dictionary with registration number 3.

(5) Output the fifth character b with code 2, increment the registration number of b in the dictionary by one, and decrement the registration number of a by one. (6) Output the sixth character c with the code 3, increment the number of c in the dictionary by one, and decrement a by one. (7) Output the seventh character a with code 3, increment a's number in the dictionary by one, and decrement c by one.

[0023] Similarly, the characters that appear are encoded using the registration numbers in the dictionary, and the registration numbers of the characters that appear are incremented by one and registered. According to the above-mentioned self-organization method, the more frequently appearing characters are rearranged, the smaller the registration number will be, and the smaller the registration number, the smaller the number of bits it will represent, which will reduce the amount of data in the output code. can do. In addition, in the SOR encoding method, the more frequently appearing characters are placed closer to the beginning of the dictionary, so
When the SOR code is arithmetic encoded, it is possible to reduce the burden of sorting the appearance frequencies in the arithmetic encoding.

FIG. 1 shows the basic configuration of the present invention. In the figure, 1 is an SOR encoding unit, 2 is a multi-level arithmetic encoding unit, 3 is a dictionary in which character strings are registered in correspondence with registration numbers, and 4
is a counter that counts cumulative frequency. Reference numeral 5 designates a search/registration unit that searches a dictionary for input character strings and registers newly appearing character strings in the dictionary, and 6 a dictionary rearrangement unit that rearranges the registration numbers of character strings in the dictionary.

[0025]

[Operation] The operation of the configuration shown in FIG. 1 will be explained. The character string input to the SOR encoding unit 1 is referred to the dictionary 3 by the search/registration unit 5. If the same character string exists in the dictionary, its registration number is output as an SOR code. Furthermore, if the input character string is unregistered, the search/registration unit 5 registers the character string in the dictionary 3 and outputs the input character string as raw data. Furthermore, the dictionary sorting unit 6 is configured to use MTR or T
The dictionary is updated according to self-organization rules such as R-way, and the registration numbers of character strings in the dictionary are rearranged.

The SOR code output from the SOR encoder 1 is input to the multilevel arithmetic encoder 2. On the other hand, the counter 4 counts the appearance frequency and cumulative frequency of each character string, and inputs the count to the multilevel arithmetic encoding unit 2. Multilevel arithmetic encoding unit 2
calculates the interval width, the upper and lower ends of the interval corresponding to the input character string based on the input SOR code, the number of the character string, the cumulative appearance frequency, etc., and calculates any numerical value between the intervals. Output as sign.

[0027]

[Embodiment] An embodiment of the present invention will be explained with reference to FIGS. 2 to 6. In the present invention, in SOR encoding, the registration number of a character string obtained by referring to a dictionary is used as a variable length code in which the smaller the registration number, the shorter the code length. FIG. 2 shows a variable length encoding method for registration numbers according to the present invention. Figure (a) shows an example of a variable length code. Assuming that the appearance probabilities of symbols a, b, c, and d are 0.5, 0.25, 0.15, and 0.1,
Code “1” for a, code “01” for b
, "001" for c and "000" for d, short codes are used for symbols that appear frequently.
The data compression rate can be improved by encoding symbols with a variable length code such that long codes are assigned to symbols that appear less frequently.

As an example of a variable length code, Elias (El
ias) There is a code. Figure (b) shows Elias (El
ias) indicates the code. Elias' gamma code is a binary number that represents a number with a prefix that represents the number of digits, and the prefix adds a zero equal to the number of digits in the binary number minus one. . For example, as shown in the figure, the number 1 is the code "1" without any prefix, and the number 2 is 2
As a prefix to the binary number "10", add one digit (number of digits 2-1 =
1) Add 0 in front of "10". Number 3 is a binary number ``
11'' and the number of digits is 2, so add one digit 0 as a prefix. Similarly, the number 4 is the binary number "100" and has three digits, so two digits of 0 are added as a prefix.

According to such an alias variable length code, it is possible to uniquely decode the code word even if the code word is packed with bits. Elias' δ code is encoded using Elias' γ code. In Elias' δ code, the γ code is used as the upper bit, and lower bits equal to the number subtracted by 1 from the number in the gamma code are added. Then, the binary representation of the lower bits represents the number starting from the number of the upper bits. For example, "010" indicating number 2 of the γ code is used to represent numbers 2 and 3. The upper bit of each is "010", and subtracting 1 from number 2 is 1, so by adding 1 bit "0" to "010", number 2 and 1 bit "
The number 3 is represented by adding "1" to "010". Similarly, numbers 4, 5, 6, and 7 use "011" of number 3 in Elias's γ code as the upper bit, and the number of bits to be added is 3 - 1 = 2 bits, and "
By adding "00", "01", "10", and "11", number 4 becomes "01100" and number 5 becomes "0110".
1”, number 6 is “01110”, number 7 is “0111”
Expressed as Similarly, as shown in the figure, the information is represented by variable length codes.

FIG. 3 shows the flow of the SOR code that is encoded before arithmetic encoding in the present invention. S1 (1) Initialize the dictionary. D(i) = 0 (i = 1, 2, ... A) However, A is the size of the alphabet (number of alphabets, symbols, etc.)
maximum of 256). (2) Assign a number to each character. I(i) = 0 (i = 1, 2, ..., A) (3) Number of dictionary slots used n = 0 S2 Input one character k. S3 Identify number j of character k (j=I(k)). S4 Determine whether number j of character k is 0. If j=0, the character k is not registered in the dictionary, so the process advances to S5. If j=0, the process advances to S7. S5 If j=0, set n=n+1, increment n by 1, and encode number n with Elias code. Encode raw data k. S6 Register character k with number n in the dictionary (D(n)=
k, I(k)=n). S7 If j is not 0 in S3, character k has already been registered in the dictionary with number j, so number j is encoded with an alias code.

[0031] S8 Update the dictionary using MTF or TR, return to S2, and input the next character k (MTF
follows S9, and TR follows S10). S9 This is the MTF processing in S8. Decrease the number I(s) of the characters from s=j to 2 and the registration number D(s) by one (I(D(s))=I(D(s-1)
). Then, register the character k with the number 1 in the dictionary (I(1
)=k, D(k)=1). S10 This is the TR processing in S8. If j is not 1 (j!=1), replace the number of the character k and the registration number in the dictionary with the character number and registration number of the previous registration number in the dictionary (r = D (j - 1), D (j-1)
=k, D(j)=r, I(k)=j-1, I(r)=j
). In the present invention, S of the input character string obtained in the above flow is
Arithmetic encode the OR code.

FIGS. 4 to 6 show the flow of arithmetic encoding according to the present invention. FIG. 4 shows a flow when the history of appearing characters is not considered. S1 (1) Allocate all single characters to dictionary D. D(i)=i(i=1,2,...,A), where A
is the size of the alphabet (the number of alphabets and symbols, etc.), which is 256 at maximum. (2) Assign a number to each character. I(i)=i(i=1,2,...,A). (3) Initialize the appearance frequency (app(i)) of each character. app(i)=0(i=1,2,...,A)(4)
Initialize the cumulative frequency (cum freq(i)). cum freq(i)=A−i(i=1,...
・,A). (5) Number of dictionary slots used n=0.

S2 Input one character k. S3 Find the number of character k and perform arithmetic encoding (the flow of arithmetic encoding is the same as that in FIG. 9, so the explanation will be omitted). S4 Update dictionary by MTF or TR (MT
F is shown in S6, TR is shown in S7). S5 Increment the cumulative frequency of characters with numbers smaller than j by 1 (j=j-1, cum in the range j>0)
freq(j)=cum freq(j)+1). Then, the processing from S2 onwards is repeated. S6 MTF processing. Since it is the same as MT in FIG. 3, the explanation will be omitted. S7 TR processing. If the appearance frequency of the j-th character is 0 ((app(j)=0), app(j)=1
and n=n+1. Let registration number D(n)=r,
Replace the registration number in the dictionary (D(j-1)
=k, D(j)=r, I(k)=j-1, I(r)=j
). If app(j) is not 0, the registration number in the dictionary is replaced by setting r=D(j-1).

FIG. 5 shows the flow of arithmetic coding according to the present invention, taking single history into account and considering combinations of two consecutive characters. The diagram shows the structure of the dictionary.
A dictionary is created for each character, and a dictionary is created for each character following each character. Therefore, D(p,i)=i(p,i=1,2,...A),
Let i(p,i)=i(p,i=1,2,...,A). Regarding the appearance frequency and cumulative frequency, the case where only the first character of two consecutive characters is counted is shown (app(i), cum freq(i
), (i = 1, 2, ..., A), the appearance frequency and cumulative frequency are determined using only the number i). The rest of the flow is the same as in the case of FIG. 4 in which the history is not taken into account, so the explanation will be omitted.

FIG. 6 shows a case in which a single history is taken into consideration in arithmetic coding, and a single history is also taken into consideration for appearance frequency and cumulative appearance frequency. The flow in the figure is as follows: The appearance frequency is I(p, i), and the cumulative appearance frequency is cum freq(
p, i) (i=1, 2, . . . , A) is the same as the flow shown in FIG. 4, so the explanation will be omitted.

[0036]

[Effects of the Invention] According to the present invention, the data input to the arithmetic encoding unit is inputted data that has been sorted approximately in the order of appearance frequency, so the process of counting the appearance frequency in arithmetic encoding is simplified. , it becomes possible to improve the processing speed of arithmetic encoding.

[Brief explanation of the drawing]

FIG. 1 is a diagram showing the basic configuration of the present invention.

FIG. 2 is a diagram showing a variable length encoding method for registration numbers.

FIG. 3 is a diagram showing the flow of SOR encoding.

FIG. 4 is a diagram showing the encoding method (1) of the present invention (when there is no history).

FIG. 5 is a diagram showing the encoding method (2) (in the case of single history) of the present invention.

FIG. 6 is a diagram showing the encoding method (3) (in the case of single history) of the present invention.

FIG. 7 is a diagram illustrating the principle of a multi-level arithmetic code.

FIG. 8 is a diagram showing the configuration of a conventional arithmetic encoding device.

FIG. 9 is a diagram showing a flow of conventional multilevel arithmetic coding (in the case of single history).

FIG. 10 is a diagram showing a flow of conventional multilevel arithmetic coding (in the case of double history).

FIG. 11 is a diagram showing encoding based on self-organizing rules (SOR encoding).

[Explanation of symbols]

1: SOR encoding unit, 2: Multi-level arithmetic coding unit, 3: Dictionary, 4: Counter, 5: Search and registration department, 6: Dictionary sorting section

Claims (4)

[Claims] [Claim 1] In a method of encoding input character strings using arithmetic codes and compressing and decompressing the data, each time a character string is input, the character strings are rearranged so that the more frequently appearing character strings, the lower the registration number in the dictionary. , an SOR encoding unit (1
) and the SOR output from the SOR encoder (1)
an arithmetic encoding unit (2) that inputs a code and performs arithmetic encoding; and after encoding the input character string with the SOR code,
A data compression method characterized in that the code is encoded using an arithmetic code. [Claim 2] The arithmetic code according to claim 1 is processed by SOR.
A compressed data restoration method characterized by converting into a code and then restoring a character string based on the SOR code. [Claim 3] In claim 1, the registration number of the input character string in the dictionary is incremented by 1 toward the smaller value, and
A data compression method that decrements the registration number of the character string that was incremented by one. [Claim 4] In claim 1, the registration position of the input character string in the dictionary is moved to the beginning of the dictionary, and the character string registered from the beginning of the dictionary to before the registration position where the character string was registered is A data compression method characterized by decrementing each registration number by 1.