JPS6356020A - Compression decoding device - Google Patents

Abstract

PURPOSE:To decode a signal efficiently without requiring much table memory capacity by subtracting a predicting value calculated by a predicting device from a signal to obtain a predicting error and referencing the table memory based on a correction value of a signal being the division of the predicting error by the standard deviation of the signal. CONSTITUTION:An arithmetic decoder 103 executes arithmetic decoding to a signal compressed based on the appearance probability and accumulated appearance probability of a signal fed from the table memory 102 to decode the signal. The signal is fed also to a predicting device 106, the predicting value and its standard deviation of the inputted signal are predicted from the inputted signal in the past and the result is outputted. A subtracter 105 subtracts the predicting value outputted from the predicting device 106 from the signal and outputs the result. A divider 104 divides the output of the subtracter 105 by the standard deviation outputted from the predicting device 106, gives the result to an address signal terminal of the table memory 102 and the table memory 102 gives the probability recorded in the address represented by the output of the divider 104 to the arithmetic decoder 103.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention provides a method for removing redundant noises contained in a signal from a compressed signal in order to save transmission time and storage capacity during signal transmission or recording. , relates to a compression/decoding device for restoring an original signal.

[Conventional technology]

2. Description of the Related Art Conventionally known and used methods for compressing redundant data and restoring original redundant data from compressed data include a method using a code called an arithmetic code. For details on arithmetic code data compressors (hereinafter simply referred to as arithmetic encoders) and data decoders (hereinafter simply referred to as arithmetic decoders), see, for example, International Business Machines Corporation of the United States.
s Machines Corporation) to 1
The journal IBM Journal of Research and Development published in 1976
BM Joroal or Re5eacb andD
eve1op+5eot) Volume 20 No. 3 198-2
03 page, pages 135-149 of Vol. 28, No. 2 of the same magazine published in 1984, and Stanford University in the United States (
5tanford University) in 1976
Richard Clark Bust (Ric
doctoral dissertation by HardClark Pa5co) [Source Coding Algorithm for First Data Compression (Source codi
ng algorithm for fast data
compression)”.

Of these methods, the method for compressing data can be briefly described by compressing a signal sequence, that is, an information digit sequence
, X2, . . . , xN are converted into code digit strings Wl, W3, . be removed.

First, let F be the value at the smaller end of a predetermined interval on the number line (0 or more and less than 1) (i.e., 0), and T
Let be the width of the interval (i.e. 1). Next, the following steps AI) to A5) are executed. Note that in the following, X←y means assigning the value of y to X. It is also assumed that all operations are performed in the b-adic system.

AI) Set the value of i to 1.

A2) Input the i-th information digit X1.

A3) The probability of appearance of the i-th information digit XI (1+
(X +) and cumulative appearance probability Ci (x +) − Ey < xc
Find z(y).

A4) F = F + Ci(x +) ・TT4-
Let q + < x 1) - T, and truncate T to a predetermined significant figure.

A5) i (If N, increase the value of i by 1 and move to A2).

If i=N, end.

Finally, among the real numbers included in the interval determined by the final values of F and T, choose a real number whose digit I when expressed numerically is small, and realize the realized digits O, w1w of the selected real number.
2...The part below the decimal point of WL is the sign digit string W
l, VV2,..., WL. It is shown in the above-mentioned document that if the above steps (1) to A5) are followed, the length of the code word is generally smaller than the length N of the information digit string, and data can be compressed.

On the other hand, the method for restoring data will be briefly described as follows: code Digi-Soto sequence VVI, w2, . . .

For example, VVL is converted as follows, and the original information digit string Xi, X2, . . . , xN is restored.

First, 01w1w2...wL is assigned to W, and 1 is assigned to S. Next, execute steps 1 [B]) to B5) below. In the following, it is assumed that X 4 - y means assigning the value of y to X. It is also assumed that all operations are performed in the b-adic system.

Bl) Set the value of i to 1.

B2) Let XI be the largest digit among y that satisfies Ci(y)≦W/S.

B3) Output x as the recovered information digit.

B4) W=W C1(x+)・S S　”　q　+(X　+)・S and truncate S to a predetermined significant figure.

B5) If i<N, increase the value of i by 1 and move to B2)/\\.

If i=N, end.

If you follow the above steps Bl) to B5), the code digit string V
The above-mentioned document shows that the original information digit string xl, x2, . . . , XN can be reconstructed from VH, VV 2 , .

Arithmetic coding is an encoding method that efficiently compresses signals as long as the probability of occurrence of the signal is given, and the problem is how to obtain the probability of occurrence. For this purpose, a method is generally used in which the appearance frequency of a signal is recorded in a memory in parallel with the encoding/decoding process, and the appearance probability is calculated.

Now, an arithmetic encoder and an arithmetic decoder for compressing data and restoring data as described above are described in, for example, U.S. Patent No. 4,122.440 of International Business Machines Corporation of the United States. This can be realized with a circuit that includes arithmetic operation circuits such as addition and multiplication, as shown in the figure below.

[Problem that the invention seeks to solve]

However, in the conventional method, when the bit length of the signal is about 1 bit and the number of levels is small, the capacity of the daple memory for recording the probability of appearance of the signal may be relatively small, but the bit length of the signal becomes large. As the number of signal levels increases, there is a drawback that the table memory capacity for recording the probability of appearance of the signal becomes enormous.

However, depending on the application, it is required to compress signals with a long bit length and a large number of levels. For example, when compressing and recording quantized analog signals in audio-visual equipment for awards such as digital audio discs, the number of signal levels is increased to maintain sound and image quality. It is essential that

The present invention provides a compression decoding device compatible with a compression encoding device that efficiently compresses a signal without requiring a large amount of table memory capacity for recording appearance probabilities even if the number of signal levels is large. That is.

[Means for solving problems]

The present invention includes a first predictor that outputs a predicted value and standard deviation of an input signal based on a signal sequence input in the past, and a second predictor that subtracts the predicted value outputted from the first predictor from the signal. 1, a first divider that divides the output of the first subtractor by the standard deviation of the first predictor, and a pre-recorded probability according to the output of the first divider. and an arithmetic encoder that performs arithmetic encoding on the signal according to the output of the first table memory. A compression decoding device that decompresses the input signal from a signal decoded by the compression decoding device, the second predictor having the same function as the first predictor;
a second subtractor that subtracts the predicted value output from the predictor; a second divider that divides the output of the second subtractor by the standard deviation of the second predictor; and a pre-recorded probability. and an arithmetic decoder that performs arithmetic decoding in accordance with the output of the second table memory. .

1. Effects Arithmetic coding is an encoding method that efficiently compresses signals as long as the probability of occurrence of the signal is given, so the problem is how to obtain the probability of occurrence. Conventionally, the frequency of occurrence of a signal is recorded in memory. However, the probability structure of the signals around us is usually known to some extent. For example, in the case of speech, the distribution is close to the Gaussian distribution, and in the case of English, the alphabet [[
It can be seen that many eJs appear. If the probability distribution of a signal is known in advance in this way, the probability of occurrence of the signal can be calculated by calculating the probability of occurrence using an approximation formula for the known distribution, or by storing the probability of occurrence of the signal in a table memory in advance. One possible method is to record it in a table and refer to it when necessary.

Among these, the method using approximation formula 3 is, for example, if the signal is generated according to a normal distribution, 198
Iwanami Mathematics Dictionary 2nd Edition 12th Printing No. 979 published in 0
The approximation formula for the normal distribution as described on the page is,
This method calculates the value by substituting the value of the signal.

The method of referring to the signal occurrence probability table recorded in advance in the table memory is to calculate the occurrence probabilities of all possible signals in advance using the approximation formula or the revised version by Marcel Pol published by Hakusuisha in 1974. Expanded Universal Numerical Table 8th Printing 5
The probability of occurrence can be determined based on numerical tables and formulas such as those described on pages 40 to 686, or by measuring the frequency of occurrence of the signal, and the obtained probability of occurrence is recorded in the table memory as an occurrence probability table. , this table memory is referred to when calculating the probability of signal occurrence.

However, although the method using formulas yields relatively accurate occurrence probabilities, it takes time to calculate, so when you need to calculate occurrence probabilities quickly, refer to a signal occurrence probability table recorded in table memory in advance. The method is valid. Furthermore, by effectively utilizing the table memory, it is possible to compress signals with a large number of levels.

However, there are problems with using table memory. In other words, even if the type of distribution is known, if its mean or standard deviation is unknown or fluctuates, you should rewrite the contents of the table memory each time the mean or standard deviation changes, or write all the data in advance. Probability tables for possible means and standard deviations should be prepared. Therefore, in the present invention, the prediction error is obtained by subtracting the average value of the signal (that is, the predicted value calculated by the predictor) from the signal, and the prediction error is further divided by the standard deviation of the signal. to refer to table memory. In this way, the average and standard deviation of the finally obtained correction value will always remain constant even if the signal average and standard deviation change, so the appearance probability of the correction value, not the signal appearance probability, can be stored in the table memory. If you record it in , and refer to the table memory based on the correction value,
Even if the average or standard deviation of the signal changes, there is no need to rewrite the table memory. Actually, let the signal be X, and the average of the signal be
Letting the standard deviation of the signal be B, calculate the average and standard deviation of the signal correction value T. Since T = (X-A), /B, the average and standard deviation of T are as follows. It can be seen that it is constant regardless of the average or standard deviation of the signal.

(average of T) - ((average of X) - to 4) 7/B=
0 (standard deviation of T) - f (average of (TA)2) = vr
(Average of (x-A)2), / 13 = (standard deviation of This is determined by a predictor.

Also, what kind of predictor do you use specifically?

The presence or absence of a signal must be determined depending on the characteristics of the signal, but for example, predictors for signals such as audio and images can be found in the book by Takashi Arimoto [Signal/Image Symptoms] published by Sangyo Tosho Co., Ltd. in 1981. This is detailed in documents such as the Digital Processing Co., Ltd.

If the method described above is used, only the probability table when the mean is 0 and the standard deviation is 1 needs to be recorded in the table memory as 1-, so the capacity of the table memory can be saved.

Now, an arithmetic encoder compresses a signal using the probabilities generated by this method, but in order to restore the original signal from the compressed signal by an arithmetic decoder, the input to the arithmetic encoder is The same probabilities must be given to the arithmetic decoder. Therefore, in the present invention, the compression decoding device is equipped with the same predictor as the compression encoding device, and the signal restored by the compression decoding device is input to the predictor included in the compression encoding device. do. l j of the predictor
Tl-like! If the initial state of the predictor of the compression encoding device is made the same as the initial state of the predictor, the subsequent transition of the internal state of the predictor will match the transition of the predictor of the compression encoding device. It becomes possible to calculate the same probability as the probability given to the arithmetic encoder from the predicted value to be output. Specifically, similar to the compression encoding device, the prediction error is obtained by subtracting the average value of the signal (that is, the predicted value calculated by the predictor) from the signal, and then the prediction error is divided by the standard deviation of the signal. The probability of occurrence of the signal is obtained by referring to the table memory based on the correction value of the signal. Note that the contents of the table memory are the same as those of the table memory included in the compression encoding device.

FIG. 1 shows a basic configuration diagram of the present invention. In the figure, the compressed signal is input from an input terminal 101. The compressed signal is supplied to an arithmetic decoder 103, and the arithmetic decoder 103 performs arithmetic decoding on the compressed signal based on the probability of occurrence and cumulative probability of occurrence of the signal supplied from the table memory 102. and restore the signal. The restored signals are supplied to the output terminal 107 and sequentially output. On the other hand, the signal is also supplied to the predictor 106, which predicts the predicted value of the input signal and its standard deviation value based on the previously input signals, and outputs the result. The subtracter 105 subtracts the predicted value output from the predictor 106 from the signal and outputs the result. Subtractor 104
divides the output of the subtracter 105 by the standard deviation value output from the predictor 106, and supplies the result to the address signal terminal of the table memory 102. The table memory 102 supplies the probability recorded at the address indicated by the output of the divider 104 to the arithmetic decoder 103.

Note that FIG. 2 shows a compression encoder to which the compression decoder of the present invention corresponds. In the figure, a signal is input from an input terminal 201. The signal is supplied to the predictor 202, and the predictor 202 predicts the predicted value of the input signal and its standard deviation value from the previously input signals, and outputs the result.

The subtracter 203 subtracts the predicted value output from the predictor 202 from the signal and outputs the result. A divider 204 divides the output of the subtracter 203 by the standard deviation value output from the predictor 202 and supplies the result to the address signal terminal of the table memory 205. Table memory 205 supplies the probability recorded at the address indicated by the output of divider 204 to arithmetic encoder 206 . The arithmetic encoder 206 inputs the input terminal 201 according to the probability supplied by the table memory 205.
Arithmetic encoding is performed on the signal supplied from the , and the code is supplied to the output terminal 207 . The codes are sequentially output from the output terminal 207 and go 9.

〔Example〕

FIG. 3 shows an embodiment of the present invention. In the figure, blocks having the same functions as in FIG. 1 are given the same numbers. In the figure, the compressed signal is input from an input terminal 101. The compressed signal is supplied to an arithmetic decoder 103, and the arithmetic decoder 103 is supplied with a table memory 10.
Arithmetic decoding is performed on the compressed signal based on the appearance probability and cumulative appearance probability of the signal supplied from 2 to restore the signal. The restored signals are supplied to the output terminal 107 and sequentially output. On the other hand, the signal is
6, predicts the predicted value of the input signal and its standard deviation value from the previously input signal, and outputs the result. Inside the predictor, delay circuits 301, 302, and 303 are connected in series, and signals up to three times ago are held. The signals held in the delay circuits 301, 302 and 303 have constant v1. al
, 32, and 33, and the sum of the multiplication results is calculated by an adder 307 and used as a predicted value of the signal. on the other hand,
In this embodiment, in order to obtain the standard deviation value of the predictor, the subtracter 308 takes the difference between the signal and the predicted value, that is, the child side error, the multiplier 30 calculates the square of the result, and the integrator 310 calculates the square of the result.
The output of the multiplier 309 is integrated, and the square of the refined result is taken by the squarer 311 to obtain the standard deviation value. However, in this embodiment, the integrator 310 does not perform strict integration, but calculates the average value of the outputs of the multiplier 309 over the past three times. A subtracter 105 subtracts the predicted value output from the predictor 106 from the signal and outputs the result. A divider 104 divides the output of the subtracter 105 by the target 1 deviation value output from the predictor 106. and supplies the result to the address signal terminal of the double memory 102.

However, the output of the divider 104 is rounded to a binary representation of 12 significant digits and output. table memory 10
2 is 2 ROMs (read only memory > 32
It consists of 1,322 pieces. Each ROM has a 12-bit address signal terminal, an 8-bit word length, and a capacity of 4096 bytes. Divider 1 is in ROM321
The probability of appearance of the correction value of the signal output by 04 is determined by the ROM32
2, the cumulative appearance probabilities are recorded in a binary representation of 8 bits to 8, respectively, and the probabilities recorded at the addresses indicated by the output of the divider 104 are respectively supplied to the arithmetic decoder 103.

FIG. 4 shows an example of a compression encoder to which the compression decoder of FIG. 3 corresponds. In FIG. 1, blocks having the same functions as those in FIG. 2 are given the same numbers. In FIG. The signal is supplied to the predictor 202, and the predictor 202 predicts the individual values of the input signal and its standard deviation value from the signals input in the past, and outputs the results. Inside the predictor, delay circuits 401, 402, and 403 are connected in series, and signals up to three times ago are held.

The signals held in the delay circuits 401, 402, and 403 are multiplied by predetermined constants al and a2 and a3 in constant multipliers 404, 405, and 406, respectively, and the sum of the multiplication results is added in an adder 407. It is calculated and used as the predicted value of the signal. On the other hand, in this example, in order to obtain the standard deviation value of the predictor, the subtracter 408 takes the difference between the signal and the predicted value, that is, the prediction error, and the multiplier 409 calculates the square of the result.
An integrator 410 integrates the output of the multiplier 409, and a squarer 411 measures the parallelism of the integration results to obtain a standard deviation value. However, in this embodiment, the integrator 410 does not perform strict integration, but calculates the average value of the outputs of the past three times among the outputs of the multiplier 409. Subtractor 203
subtracts the predicted value output from the predictor 202 from the signal and outputs the result. The divider 204 divides the output of the subtracter 203 by the standard deviation value output by the predictor 202,
The result is supplied to the address signal terminal of table memory 205. However, the output of the divider 204 is 12 significant figures.
The output is rounded to a binary representation of bits. The table memory 205 consists of two ROMs (ROM<read-only memory) 421 and 422. Each ROM is
The address signal terminal is 12 bits, the word length is 8 bits, and the capacity is 4096 bytes. In R,0M421, the probability of appearance of the correction value of the signal output by the divider 204 is
In the ROM 422, each cumulative appearance probability is recorded in 8-bit binary representation. The arithmetic encoder 206 inputs the input terminal 201 according to the probability supplied by the table memory 205.
Arithmetic encoding is performed on the signal supplied from the , and the code is supplied to the output terminal 207 . The codes are sequentially output from the output terminal 207.

〔Effect of the invention〕

As described above, according to the present invention, even if the number of levels of a signal is large, a compression encoding device that efficiently compresses a signal without requiring a large amount of table memory capacity for recording appearance probabilities can be realized. A corresponding compression/decoding device can be easily configured.

Therefore, it is clear that the present invention is effective in fields where compression of signals with long bit lengths is required, such as data compression in audiovisual equipment for viewing purposes.

[Brief explanation of drawings]

FIG. 1 is a basic configuration diagram of the present invention, FIG. 2 is a basic configuration diagram of a compression encoding device to which the compression decoding device of FIG. 1 corresponds, and FIG.
The figure is a block diagram showing one embodiment of the present invention, and FIG.
FIG. 2 is a block diagram of a compression encoder to which the compression decoding device shown in the figure corresponds. 101.201...input terminal, 106,202...
Predictor, 105,203...Subtractor, 104.207
1... Divider, 102, 205... Table memory, 106... Arithmetic encoder, 107, 207... Output terminal 9

Claims (1)

[Claims] a first predictor that outputs a predicted value and standard deviation of an input signal based on a signal sequence input in the past; and a first subtractor that subtracts the predicted value output from the first predictor from the signal. a first divider that divides the output of the first subtracter by the standard deviation of the first predictor; and a first divider that outputs a pre-recorded probability according to the output of the first divider. 1 table memory, and an arithmetic encoder that performs arithmetic encoding on the signal according to the output of the first table memory, and compresses the signal. A compression decoding device for restoring an input signal, the second predictor having the same function as the first predictor, and a prediction output from the second predictor from the signal decoded by the compression decoding device. a second subtractor that subtracts a value; a second divider that divides the output of the second subtractor by the standard deviation of the second predictor; and a pre-recorded probability A compression decoding device comprising: a second table memory that outputs an output according to the output of the second table memory; and an arithmetic decoder that performs arithmetic decoding according to the output of the second table memory.