JP3407588B2 - Encoding / decoding device - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to entropy coding, and more particularly to a processing procedure and a device configuration of a coding / decoding device.

[0002]

2. Description of the Related Art The process of converting a signal into a code generally comprises a modeling section and an entropy coding section. The modeling unit is intended to convert a signal sequence. For example, the image signal is divided into blocks each including a plurality of pixels and the signals in the blocks are converted into frequency components. Alternatively, the modeling unit converts the signals into a continuous number having the same signal level. There are many methods such as On the other hand, the entropy coding unit allocates codes to the signal series created by the modeling unit. The ratio between the input signal and the number of bits of the code is expressed as a compression rate. It is desired that the compression rate is high, and therefore many configurations of the modeling unit and the entropy coding unit have been proposed.

The previously proposed code generation algorithms and their characteristics and properties are described in detail in "Mathematical Mathematics of Information and Coding" (1994), published by Iwanami Shoten, by Han and Kobayashi. Information theory defines an index called "entropy" that indicates the limit of achievable compression ratio. Also, there are “optimal” and “instantaneous decodable” as technical terms that represent the property of the code. Here, the optimum means a code having the smallest average code length, and the instantaneous decodable means that each code can be decoded without looking at all code sequences. These are important properties for actually realizing the device. Conventionally, the Huffman code proposed by Huffman in 1952 is known as a code having both properties (Huffman, DA: A method fo
r theconstruction of minimum-redundancy codes, Pro
c. IRE, vol. 40, pp. 1098-1101, Sept. 1952).

As a method of using the Huffman code, the total occurrence probability of an event to be encoded is measured in advance, a code based on this occurrence probability is generated, and the code is stored in a memory in a table format. Then, the actual code target event is obtained by a table search and converted into a corresponding code. It is used in a wide range of fields because "optimal" and "instantaneous decodable" can be realized with simple procedures. Further, the Huffman code is classified as a block code because of the property that one code corresponds to one signal.

[0005]

The Huffman code is an entropy coding system that can theoretically realize "optimal" and "instantaneous decodable". However, in the actual encoding process, for example, (a) a change in the occurrence probability of the encoding target (b) a signal processing procedure of the modeling unit (c) a signal process in consideration of a device configuration or the like for realizing the above If you do not, you cannot achieve efficiency improvement.

First, with respect to (a), the occurrence probability of an object to be coded is not necessarily constant in nature, but generally fluctuates due to various factors. In the method in which the Huffman code is set in advance as the code table, the efficiency decreases when the occurrence probability of the encoding target changes. Even if an attempt is made to adaptively generate a code in response to changes in the occurrence probability of an event, the Huffman code code generation procedure becomes a tree structure in which the event with the smallest probability value is executed to the event with the highest probability value. Therefore, the code can be used only at the time when all the code generation procedures are completed. Therefore, in order to perform adaptive coding, the processing load becomes enormous and the device configuration becomes complicated. To summarize these characteristics, there is no change in the occurrence probability value of an event. In other words, Huffman codes are suitable for static coding targets, but the opposite case for dynamic coding targets. Can be said to be unsuitable.

Next, various methods have been proposed and used for the signal processing of the modeling unit of (b). As an example of performing conversion processing by collecting a plurality of signals, there is a run length used in a facsimile. In the basic MH (Modified Huffman) method, the number of consecutive identical signal levels (run length) is measured, and the result is converted into a prepared Huffman code. However, the probability of occurrence of consecutive run lengths changes depending on the origin of the runs in one line, and even though the run length is initialized to 0 for each line, a fixed code table is used to compress the run length. It is a factor that causes a decline in the rate. In this way, efficiency improvement cannot be expected unless entropy coding is performed after grasping the contents of the signal processing procedure of the modeling unit.

In order to realize the coding / decoding processing for (c), a temporary storage means for the signal to be coded is indispensable. In the conventional proposal, the viewpoint of the realization means by the semiconductor circuit is omitted, and in particular, the speedup of the data input / output of the memory is a major factor that determines the performance of the device. Not considered in combination. A dedicated LSI for executing the above encoding / decoding processing has been developed, but for example, there is no one mixed with a memory function capable of accumulating compressed data for one screen on the same chip, and it is configured on another chip. Since the combined memory is combined by wiring, the problem of execution speed remains.

[0009]

In order to achieve the above object, according to the present invention, an event that an input can take and a probability that this event can occur are stored in a memory in descending order of probability, and the stored probability and The events are read in order from the one with the highest probability, the code length of the code is determined from the read probability, the cumulative probability is calculated based on this code length, and the code is generated from the calculated code length and cumulative probability. It has a feature.

[0010]

BEST MODE FOR CARRYING OUT THE INVENTION An embodiment of the present invention will be described below with reference to the drawings.

First, the encoding device will be described.

FIG. 2 shows the basic configuration of the encoding system.
A signal input to the encoding system receives an analog signal from the outside by the input device 200, and this signal is converted into a digital signal by the A / D converter 201. A code is created by the encoding device 203 in the converted digital signal, and the code is stored in the memory 202 or output to the outside by the communication device 204 and sent to another system. It should be noted that the input device 200 may be a device such as a scanner that reads an original document in addition to simply receiving an input from the outside. Further, regarding the A / D converter 201,
If the input signal is a digital signal, there is no problem even if it is omitted.

The encoding process is intended to reduce the amount of information in a continuously input signal sequence by replacing the events that can be taken by each signal forming the sequence with a code. For example, when the input signal is composed of a 3-bit binary number, there are eight kinds of events that the signal can take, and the probability of inputting each signal is, for example, if the input device is a scanner, it is drawn on the manuscript. It depends on the content.

FIG. 1 shows a coding device 203 of the coding system.
Is shown. The management unit 101 includes a code generation unit 10
2. It manages the coding error calculation unit 103 and the probability setting unit 104, and starts and stops each unit as necessary. The probability setting unit 104 performs initial setting, associates an event that can be taken by an input signal given from the outside with a probability of occurrence, and the memory 202 in descending order of probability of occurrence.
To store. If the probability corresponding to the value that the input signal can take is not clearly known, a probability function, statistical value, etc. are given. The code generation unit 102 reads events and probabilities according to the order set in the memory 202 by the probability setting unit 104, and generates a code corresponding to the input signal. Therefore, it is possible to generate an optimal and instantly decodable code. The coding error calculation unit 103 obtains a coding error that occurs when the code length of each event is set, and uses the coding error to correct the occurrence probabilities of the remaining events for which code generation has not ended.

Encoding is based on the occurrence probability value of an event,
It is to replace with a code with a finite number of bits. Basically, a short code is assigned to an event with a high probability value, and conversely, a long code is assigned to an event with a low probability value. Here, it is important to shorten the average code length as much as possible, but it is the "optimum" code that realizes this limit.

The encoding of the present invention is based on the event k that can be input.
A code is generated from a probability value represented by an s-adic number in association with (occurrence event), and an "optimal" code with a short average code length is generated.

Next, the outline of code generation when s = 2, that is, in the case of a binary code will be described with reference to FIG.

First, initial setting is performed (301). As an initial setting, an event k _i (i = 1
~ N) and the probability P _i (i = 1 to n) of occurrence of the event K _i are associated with each other and are arranged in order from the one with the highest probability. For example, if the input signal to the encoder 203 is 3 bits, possible events are (000), (001), (01
0), (011), (100), (101), (110), (1
11), and the probability that each event occurs is P ₁ = 0.21, P ₂ = 0.19, P ₃ = 0.18, P ₄ =
0.17, P ₅ = 0.10, P ₆ = 0.10, P ₇ = 0.0
If 5, P ₈ = 0.00, the result is as shown in FIG.

Next, when an input signal is input (30
2) Obtain the bit length (code length) b _k of the code (30
3). As described above, since the present invention generates a code based on a binary probability value, b _k = log (1 / P _k ) (base is 2). Then, the obtained b _k is rounded off after the decimal point. However, when the probability value is obtained from the rounded b _k , the rounding may change the rank shown in FIG. In the present invention, since the code is generated from the one having a large probability value, it is necessary to prevent the rank from changing. Therefore, if the ranking changes, the decimal places are rounded up or down to prevent the ranking from changing.

Next, the sign error E _k is obtained, and the sign error is distributed to the event having a low probability value (304). The sign error E _k is _calculated by E _k = (1/2) ̂b _k −P _k . When the obtained code error E _k is distributed, the distribution is performed so that the order of the probability values does not change. In this way, the average code length can be shortened by obtaining and distributing the code error E _k . That is, when rounding up is performed when setting the code length and distribution is not performed, the probability of occurrence of each event is smaller than the actual value by the code error. Also, the probability value of the entire event is reduced by the accumulated value of sign errors. In other words, the probability of occurrence will be set smaller than the actual value. As described above, the number of bits (code length) of the code is b _k = log (1 /
Since it is _obtained by P _k ), the smaller P _k is, the larger b _k is and the larger the number of bits (code length) of the code is. In the present invention, since the code error E _k is calculated and distributed, the average code length can be finally shortened.

It is judged whether or not the code length of the event that actually occurred and is to be coded is generated, and the above process is repeated using the corrected probability value. When the processing of the event actually occurred is completed, the code length b _k is set and the next processing is performed (30
5).

When the code length of the event to be coded is obtained, the bit of the code is set (306). The sign bit is a cumulative probability value represented by a binary number q _k (q _k = Σ (1/2)
^ B _k-1 ) is set by cutting out the number of bits b _k after the decimal point. Note that when k = 1, q _k = 0,
The code is a sequence of b _k 0s.

Next, the code generation will be specifically described with reference to a flowchart. Here, the input signal is 3
A case will be described in which a binary signal represented by bits is used and the probability P _k for an event k _i that the input signal can take is known in advance.

FIG. 5 shows a flowchart of the entire processing. The management unit 101 described in FIG. 1 manages and controls the start and stop of each unit according to this flowchart.

First, when the encoding device 203 is activated, initialization is performed (501). The initial setting is performed by the management unit 101 activating the probability setting unit 104. Next, the management unit 101
When the input signal IN is input (502), starts the code generation unit 102 and generates a code (503). The management unit 101 outputs the code generated by the code generation unit 102 to the memory 202 or the communication device 204, and waits for the next input signal (504).

FIG. 6 shows details of the initial setting (501), which is performed by the probability setting unit 104. When the probability setting unit 104 is activated, the event k _i that can be input and the probability P _i that occurs will be fetched from outside the encoding device 203 or from the memory 202 (601). Then, the events k _i and the probabilities P _i are associated with each other in the descending order of the probabilities P _i and stored in the memory 2
Stored in 02. By thus storing in the memory, it becomes as shown in FIG. 4 described above. Here, the flag F is for determining whether or not a code has already been generated,
If the code has already been generated, F = 1.

FIG. 7 shows details of the code generation (503) for generating the code C for the input signal IN by the code generation unit 102.

When the code generation unit 102 is activated, code generation is performed in order from the highest probability P _i stored in the memory until the input signal IN and the event k _i stored in the memory match, and the input signal IN When IN matches the event k _i stored in the memory, the code C _i is output.

When the code generator 102 is activated, the memory 2
The i-th event k _i , probability P _i , code C _i , and flag F _i from 02 are read (701). Here, F _i is judged (702), and when F _i = 1 the code C _i has already been created, so it is judged whether the input signal IN and the event k _i match (713). When F _i = 0, the code C _i is not created. Therefore, log (1 / R _i ) is calculated to obtain the number of bits of the code (703), and the calculation result is rounded off to the nearest decimal point to obtain the code length b _i . Since the rounding process may change the rank of the probability P _i , the probability P is calculated again (705),
If the rank changes, rounding up or rounding down is performed (707 to 709). Code length b obtained in this way
_A sign error calculation is performed based on _i (710). The code error calculation (710) is performed by the coding error calculation unit 103 as described above. This will be described later.

Next, the calculated code length b _i and the cumulative probability value q
_A code C _i is generated from _i . When the code is generated in this way, the corresponding flag F _i is set to 1 and stored in the memory 202 together with the generated code C _i (711, 71).
2). Then, the input signal IN is compared with the event k _i (713), and if they do not match, the cumulative probability value q
_i (= Σ (1/2) ^ b _k-1 ) is calculated (7150, and the code generation process is performed for the next highest probability P _i .

In this way, the input signal IN and the event k _i
When and match, the process ends and the code at that time is output.

Next, the code error calculation (710) performed by the coding error calculation unit 103 will be described.

FIG. 8 shows the processing of the sign error calculation (710). The code length b _i and the probability P _i are obtained from the code generation unit 102, and the probability error E _i is obtained by E _i = (½) ̂b _i −R _i (801). Here, when E _i = 0, the process is terminated without performing the distribution. When the probability error E _i is generated, the probability error E _i is distributed to the events having the ranks smaller than the probability P _i . Here, the probability error is evenly distributed and compared with the probability ranked one higher so that the rank of the probability does not change (804). When the distribution to the predetermined N0 events is completed, the distributed error is stored in the memory 202 and the process is completed.

In this way, the coding device 203 sequentially generates codes for the input signal.

Here, the case where "010" is input as the input signal IN will be specifically described. In addition, the event,
The probability and flag are assumed to be stored in the memory as shown in FIG.

When the input signal IN (= 101) is input, the event k ₁ , the probability P ₁ , the code C ₁ , and the flag F ₁ are stored in the memory 20.
Read from 2. Since F ₁ = 0, code generation is performed.
First, calculate the log ₍₁ / P 1), the calculation result code length b ₁ = 2 when rounded to the nearest whole can be obtained, examined whether variations in rank occurs when the code length b ₁ = 2 Finally, the code length is determined. In this case, since the ranking does not change, it is determined that b ₁ = 2 (FIG. 25 (a)).

Next, in order to perform the sign error calculation, E ₁ =
When (1/2) ^ b ₁ −P ₁ is obtained, E ₁ = 0.04. Therefore, the events (k _{2 to} k
Distribute the error to ₇ ). In this case, even if the error is distributed, the ranking does not change, so the distribution is performed as it is (FIG. 2).
5 (c)).

Next, since it is determined that the code length b ₁ = 2, the code C ₁ is generated from the cumulative probability value. The code C ₁ is the code length b ₁
= 2, two decimal places of the cumulative probability value q ₁ are taken from the left side to generate a code C ₁ = 00. Then, since the code is generated, the flag F ₁ = 1 is set, and the code C ₁ and the flag F are set.
Store ₁ in memory. Next, the input signal and the event k ₁ are compared, but since they do not match, the same process is performed for the event K ₂ that is ranked next. By performing the processing in this way, the code C ₃ = 100 corresponding to the input signal IN
Is obtained.

FIG. 9 shows the contents stored in the memory 202 when processing is performed for all events. Further, although the relationship between the generated codes is shown in FIG. 10, the codes are uniquely determined in this way.

Next, the case where the possible event k and the probability P are given by the probability function f (x) will be described. In this case, each time the input signal IN is given, the probability for the event is obtained and the code is generated. For example, FIG.
When the probability P is given by the probability function f (x) as shown in FIG. 11, all possible events and probabilities are taken so that the input signal IN ₁ has the maximum probability as shown in FIG. Ask. Thus to generate a code C ₁ for similar input signal IN ₁ as described above on the basis of event k and the probability P obtained by the. When there is the next input signal IN ₂ ,
1 the input signal IN ₂ as shown in (c) is to maximize the probability to obtain the all events and probabilities may take to generate a code C ₁ with respect to earlier as well as the input signal IN _1.

The apparatus configuration in this case is the same as that shown in FIG. 1, and the basic processing is the same as that shown in FIG. Therefore, in the following, the description will focus on the parts that differ in processing.

When the processing starts according to FIG. 5, the management unit 101
Then, the probability setting unit 104 is activated and initial setting is performed (501).

FIG. 12 shows the processing performed by the probability setting unit 104. First, the probability function f stored externally or in a memory
(x) is taken in and the initial value IN ₀ is set (120
1, 1202). Then, next, an event k _i and a probability P _i that can be input so that the initial value IN ₀ has the maximum probability are obtained (1203). Then, the event k _i and the probability P _i are
_The data are stored in the memory in descending order of _i , and the initialization is completed.

Next, when the management section 101 receives the input signal IN, the code generation section 102 is activated to generate a code. FIG. 13 shows the processing of the code generator 102. When the input signal IN is input, the code generation unit 102 obtains an event k _i having the i-th largest probability when the input signal IN has the maximum probability (1301). Then, the i-th probability P _i , code C _i , and flag F _i are read from the memory (1302). After reading in this way, to generate a code C _i performs the same processing as the processing described (702-712) in FIG. 7 (1,303 to 1,313) and it is judged whether the input signal IN and event k _i matches If they do not match, the cumulative probability q _i is obtained, and the code generation process is performed again. If they match, the code C _i is output.

The code error calculation 13 in the code generation is performed.
For 09, the code error calculation unit 103 performs the processing shown in FIG.

The generation of the code for the input signal IN when the probability function f (x) is given has been described above.
By partially modifying this process, it can be applied when the probability changes. The case where the probability fluctuates is, for example, a case where a code is generated for a signal read from a document as shown in FIG. 14, and a case where a probability for an event that can be input is created based on a signal input so far. . In this case, one pixel is 0 to 0 using FIG.
The case of taking 8 gradations of 7 will be described as an example. When a pixel signal is input as shown in FIG. 15A, the probability is calculated based on the input signals up to the 19th in order to generate a code for the 20th input signal. Like Then, the events K _i that the input signal can take are assigned in order from the one with the highest probability as shown in FIG. That is, when a code is generated with the i-th pixel as the input signal IN _i , the input signal (I
N ₀ , IN ₁ , IN ₂ ...... IN _i-3 , IN _i-2 , IN _i-1 )
From this, the probability P for the event K that can be input is obtained. In this case, the configuration of the encoding device is the same as that of FIG.

FIG. 16 shows the processing of the management unit 101. The management unit 101 first performs the initial setting 1601 and then the input signal I
When N is input, code generation is performed by code generation 1. Then, after inputting a predetermined number NR of input signals,
Code generation is performed by the code generation unit 102. This is because if the number of input signals IN is small, the correct probability cannot be obtained, and therefore the process is performed based on the preset probability until there is a certain number of input signals.

The probability setting unit 104 is activated by the management unit 101 to perform the initial setting shown in FIG.

When the input signal IN is input, the management unit 101
Activates the code generator 102 and generates a code based on the process shown in FIG. Next, when there is a predetermined number of inputs, the management unit 101 instructs the code generation unit 102 to switch the processing, and the code generation unit 102 executes the processing shown in FIG.

The processing of FIG. 17 will be described.

The code generator 102 obtains the probabilities P _i based on the input signals input so far, arranges the probabilities P _i in descending order of probability, and stores them in the memory 202 (1701). Then, the i-th probability P _i and the event K _i are read from the memory (1702). Unlike the above description, since the probability varies each time the input signal is input to obtain the probability with the input signal IN, the flag and the code are not stored. The subsequent processing is the processing shown in FIG.
The same processing as that of 710) is performed (1703 to 1710). Next, the input signal IN and the event k _i are compared, and if they match, a code is generated C _i . Also in this case, since the probability P _i changes, the code C _i is finally obtained.

Regarding the code error calculation, the processing of FIG. 8 already described is executed by the coding error calculation section 203.

Next, the decoding device will be described.

FIG. 18 shows the basic configuration of the decoding system. The signal input to the decoding system is a code signal and is input from the communication device 1804, and the decoding device 1802.
And the input signal is reproduced.

FIG. 19 shows a decoding device 18 of the decoding system.
02 is shown. The management unit 1901 uses the decryption unit 1
The control unit 902 manages the coding error calculation unit 103 and the probability setting unit 104, and starts and stops each unit as necessary. The decoding process of reproducing the input signal from the code performs the same process as the coding process. This is because the code is generated until it matches the input code signal, and the event corresponding to the code generated at the time of the match is output.

Then, referring to FIG. 20, when s = 2,
That is, the outline of the case where a binary code is input, decoded, and the input signal is reproduced will be described.

First, initial setting is performed (2001). For this initial setting, the same processing as the processing 301 described in FIG. 3 is performed.
Next, when a code is input (2002), the bit length (code length) b _{k of} the code and the code error E _k are obtained, and the code error is distributed to the event having a low probability value (2003, 200
4). Then, the input code CIN is compared with the generated i-th code C _i, and if they match, the event k _i is output as an input signal, and if they do not match, code generation is performed again (2006 ). The process of obtaining the code bit length (code length) b _i and the code error E _i and distributing the code error to the event having a low probability value is the process 30 of FIG.
This is the same processing as 3,304.

Next, the decoding process for reproducing the input signal from the input code will be described in detail.

FIG. 21 shows the processing of the management section 1901. The management unit 1901 activates the probability setting unit 104 and performs initial setting. When the code CIN is input (2102), the decoding unit 1902 is activated to extract the event k corresponding to the input code CIN (2103) and output it as a reproduced input signal (2104).

The probability setting unit 1904 activated by the management unit 1901 performs the same processing as that shown in FIG. That is, the event k that can be taken by the input signal in the descending order of probability P _i
Store _i and probability P _i in memory.

The management unit 1901 monitors the input code, and when the code is input, activates the input signal generation unit 1901. The cutout of the code input in the management unit can be cut out because the code is generated as shown in FIG. The decryption unit 1902 activated by the management unit 1901 is activated and the processing shown in FIG. 22 is performed. First, from the memory 1803, the i-th event k _i , probability P
_i , code C _i , and flag F _i are read (2201). Here, F _i is determined (2202), and when F _i = 1 the code C _i has already been created, so it is determined whether the input code CIN and the code C _i match (2212). F _i =
When it is 0, the code C _i is not created. Therefore, log (1 / P _i ) is calculated to find the number of bits of the code (2203), and the calculation result is rounded off to the nearest decimal point to obtain the code length b _i . Since the order of the probabilities P _i may change due to this rounding processing, the probabilities are calculated again (2205), and if the order changes, rounding up or rounding down is performed (2207 to 220).
9). A code error calculation is performed based on the code length b _i thus obtained (2210).

Next, the calculated code length b _i and the cumulative probability value q _i
To generate the code C _i . When the code is generated in this way, the corresponding flag F _i = 1 is set, and the generated code C
_It is stored in the memory 1803 together with _i . Then, the next input code CIN and code C _i are compared (221
2) If they do not match, the cumulative probability value q _i is calculated (2214), and the code generation process is performed for the next highest probability P _i .

In this way, the processing is terminated when the input code CIN and the code C _i match, and the event K _i corresponding to the code C _i at that time is output.

Next, a case will be described in which the possible event k and the probability P are given by the probability function f (x) and the generated code is decoded.

Basically, it is almost the same as the processing described above. That is, for the initial setting, the probability setting unit 104 performs the processing shown in FIG. And the management unit 1901
The decoding unit 1902 activated by performing the processing shown in FIG.

As shown in FIG. 23, the input code CI
When N is the maximum probability, the event k _i having the i-th largest probability is obtained (2301). Then, the i-th probability P _i , code C _i , and flag F _i are read from the memory (2302). After reading in this way, the same processing as the processing (2202 to 2215) described in FIG. 22 is performed to generate a code (2303 to 2316).

Next, the case where the probability changes and the generated code is decoded will be described.

FIG. 24 shows the processing of the management unit 1901. The management unit 1901 first performs the initial setting 2401, and the code CIN
Is input, the input signal is reproduced by the code 1. Then, after a predetermined number of codes NR have been input, the input signal is reproduced with code 2.

The probability setting section 104 is activated by the management section 1901 to perform the initial setting shown in FIG.

When the code CIN is input, the management unit 1901
Activates the decoding unit 1902 to generate the code shown in FIG. Next, when there is a predetermined number of inputs, the management unit 1901 instructs the decoding unit 1902 to switch the processing, and the decoding unit 1902 executes the processing shown in FIG. 25.

The processing of FIG. 25 will be described.

The decoding unit 1902 receives the probability P for the event k _i that can be used as an input signal based on the code input so far.
_i is calculated and arranged in descending order of probability.
It is stored in 1803 (2500). The process after this is shown in FIG.
The same processing (2501 to 2513) as the processing shown in (2203 to 2215) is performed.

As for the processing described above, the processing of the code generation section 102 and the processing handled by the decoding section 1902 are naturally the same.

FIG. 27 shows an encoding / decoding device in which the encoding device and the decoding device are integrated. The processing of each unit is the same as that described above. The processing described above is an example, and the present invention is not limited to this. Therefore, other processing will be described below.

(1) The generation of code bits can be simplified by replacing the occurrence probabilities of events below a certain probability value or below a certain rank with uniform probability values. For example, the signal consists of 8 bits, that is, the event is 25
If there are 6 types (2 to the 8th power), the prediction probability value is set only for the top 8 types of events, and the remaining 2
Uniform probability values can be assigned for the 48 types of events. However, the sum of the occurrence probabilities of all events must be 1.0 or less. This is an effective method when the load of sequentially calculating the probability value for each signal input is large, and the ordering according to the magnitude of the probability value may be executed for the upper eight events. For events with uniform probability values (9th to 256th in the above example),
Since a uniform code length can be assigned, the code between adjacent events of rank is a binary fraction represented by the code length.
The value of the least significant bit of (1/2) ^ b _k can be set to be the difference. As a result, the order of events is directly reflected in the code bit, so that the speed of encoding and decoding can be increased.

(2) In setting the code bit, the cumulative probability value q _k is used to generate a b _k bit code. Here, add a constant offset value to the cumulative probability value,
(1-q _k) may be utilized as the accumulated probability values.

(3) As a method of setting the code length b _k , log
Although (1 / P _k ) is rounded off, it may be rounded off. However, the encoding means and the decoding means must be unified. Here, in order to simplify the logarithmic calculation for obtaining the code length, the logarithmic calculation results of several sample points are collected in a table, and table search and interpolation processing are performed in combination based on the input signal. realizable. It is important to limit the calculation error that accompanies the interpolation process so that it does not become larger than the original when reconverted to a probability value.To do this, set the code length so that the minimum digit of the logarithmic calculation result is rounded up. Good.

Here, specific merits obtained by the present invention will be mentioned.

(1) From the viewpoint of compression rate, it has an optimal and instantaneously decodable property. A feature of the present invention is to minimize the coding error in each event and to supplement the coding error that has occurred with the remaining events to prevent loss of compression ratio. That is, the cumulative coding error can be set to 0 or close to 0 by distributing the coding error. In other words, the processing for canceling the coding error generated in each event is performed so that the cumulative coding error becomes 0 or approaches 0. The present invention realizes optimality and instantaneous decodability.

(2) In descending order of the predicted occurrence probability value,
Since it is sufficient to execute the code generation procedure up to the event that actually occurred, the average processing time for code generation can be shortened. The ability to generate codes at high speed means
This means that it is suitable for sequentially generating codes while coping with fluctuations in the occurrence probability of a coding target without preparing a code table in advance. In addition, for an event with a low occurrence probability, a certain code length can be assigned and code can be generated by ranking. On the other hand, when trying to perform adaptive code generation with the Huffman code, it is necessary to generate all codes sequentially because it is a tree structure that performs signal processing from the event with the smallest probability value to the event with the highest probability value. Must
Enormous processing time is required for code generation, which is not practical.

(3) It is possible to easily incorporate the improvement of the compression rate by expanding the information source. Assigning one code collectively for a plurality of events is known as expansion of an information source, and it is known as so-called Shannon's theorem that improvement of compression rate can be realized by this. However, if m signals of n-value level are combined, a code corresponding to n ^ m (n to the m-th power) types of events is required. If the Huffman code is used to sequentially generate codes, the code generation time becomes long and a huge code table is required, which is not practical. On the other hand, in the present invention, since the procedure until the code corresponding to the event actually occurred is generated, the signal processing can be performed at high speed and the code table is not required.

(4) The code generated by the present invention is
When viewed as numbers below the decimal point, it has the property of increasing in size according to the order of occurrence probabilities. By utilizing this property, it is possible to distinguish the codes.

Next, as an application example of the encoding / decoding apparatus of the present invention, a mixed LSI arranged on the same chip as the memory will be described.

FIG. 28 shows a mixed LSI. Mixed LS
I is an input / output control unit 280 that inputs and outputs signals to and from the outside.
1, a coding / decoding unit 28 that stores input data as a code in a memory and reproduces the data stored as a code
02, a distributor 27 that distributes data to and from a memory 2704 that stores encoded data and input data
It is composed of 03.

Each section will be described below.

The encoding / decoding section 2802 has a management section 270.
1, probability setting unit 104, code generation unit 102, decoding unit 19
02, a coding error calculation unit 103. Regarding the processing of each part, the processing when the probability is known in advance is executed (FIGS. 6 to 8 and 22).

The input / output control unit 2801 inputs / outputs data to / from an external CPU to execute activation / termination / temporary suspension, resetting, etc. of code generation / decoding processing.
Specifically, when receiving the data sent from the CPU, it notifies the management unit 2701 of the encoding / decoding unit 2802 that the data has been input. As a result, the code of the input data is generated and stored in the memory 2804. Also, C
When there is a read command from the PU, the management unit 270
The data stored in the memory as a code is read, and the decoding unit 1902 reproduces the data. Here, since the external input / output data width and the memory data width on the chip are different, the data width conversion is performed using the data input / output register 2801. As described above, since the memory access is fast and the memory data width is large, it is possible to perform control with a sufficient margin for data input / output from the external CPU. That is, high-speed data input / output can be realized with respect to the external CPU by utilizing the memory access gap by the signal processing device mounted on the chip. Therefore, based on a certain condition, it is possible to guarantee the data input / output speed that does not enter the "waiting state" for the external device while executing the internal signal processing.

The probability setting unit 104 sets a large data width that can be collectively accessed from the memory 2804, temporarily stores it in the read register 2812, and then sets the data distributor.
The 2817 is used to capture probability and event data.

The code generation unit 102, the probability setting unit 104, the decoding unit 1902, and the coding error calculation unit 103 can store the data in the middle of calculation in the memory unit without storing in the individual processing units. . Therefore, the memory 2804
And the input / output signals of each signal processing unit are collected or distributed by using the data distribution / distribution device 2811 in the signal distribution device 2803. By incorporating a memory with a large bit width, it is possible to equivalently realize high-speed memory access by inputting and outputting a plurality of types of signals with the same memory access.

The code generation unit 102 uses the read register 2812 and the data distributor 2713 to transfer the signal to be coded, which is temporarily stored in the memory 2804, to the code conversion unit and converts it into a code. Since the code generated here has a variable length, it is necessary to perform bit-wise alignment according to the data width of data input / output. Therefore, the generated code is temporarily stored in the code register 2814 and is written in the write register 2816 using the bit aligner 2815. When the write register 2816 is full, the data is transferred to the compressed data storage area of the memory 2804. Here, when the codes are arranged over two data, it is necessary to divide the code at the boundary part,
The processing procedure becomes complicated. However, if the memory data width is large, such boundary processing can be reduced, which is effective for speeding up.

The decoding unit 1902 also inputs / outputs data to / from the memory via the register, similarly to the code generation unit.

There is no limitation on the type of memory 2804 used here, and S-RAM, D-RAM, SD-RA
M, F-RAM, etc. can be used. F-RAM here
Has a characteristic that data can be stored without a power source, and is therefore suitable for storing probability, coded compressed data, and the like. With this configuration, the data input / output to / from the LSI chip is not encoded, but the data to be stored in the memory inside the LSI chip is compressed data by the encoding / decoding device 2802, so that the apparent capacity of the memory is reduced. Can be increased. Such an embedded LSI can be called a “virtual memory” because it has a different appearance from the substance of data. The compressed data stored inside the function-embedded LSI can be transferred to an external storage device or can be communicated with another device via a transmission device.

By thus performing the memory access in the LSI chip, (1) since the data width can be set large, it is possible to collectively access a plurality of data.

(2) Because of the wiring on the semiconductor chip,
Since the wiring capacity is small and the skew between the signal lines is also small,
A high transfer rate can be realized.

(3) Less susceptible to external noise and the like.

(4) The cost of the device can be reduced and the reliability is improved.

(5) Since the memory access on the chip is fast, the cache memory becomes unnecessary.

And the like can be realized.

FIG. 29 shows an example of a configuration in which a mixed virtual memory is connected to a processor for data input / output operation. Type of data to be input / output is CPU
If it is possible to distinguish by the memory address output by, the data type can be determined using the input / output control unit 2801. When the data type is data for which continuous high-speed input / output is desired, it can be converted into an actual memory address signal in the chip based on the layout configuration of the memory. For example, as one of the memory layout configurations, there is a memory block structure called a "bank", and a plurality of the banks are prepared and data can be input / output alternately to achieve high-speed memory access. The input / output control unit 2801 manages such a memory layout configuration and a memory access sequence to generate a memory address signal inside the chip, which is different from the address signal output from the external CPU, to realize high-speed memory access. it can. As another data type, when the signal is a signal for activating the encoding / decoding device, it is sufficient that the activation signal can be transmitted to the corresponding circuit without requiring the address conversion for speeding up as described above.

In order to realize this, it is effective to inform the virtual memory of the relationship between the address signal and the data type in advance from the processor. The virtual memory stores the above contact information in the LSI chip,
The nature of the signal can be judged by referring to the contents using the address signal when the access is received from the processor. Further, even without the above-mentioned notification items, the same signal type can be performed by measuring the property of the signal.
Then, a code adapted to the signal property can be generated.

In a specific device configuration, a high compression rate can be realized by combining the modeling unit and the entropy coding unit. There are two types of ideas for creating an LSI, one for general purpose and one for exclusive use.

(1) An LSI chip can be manufactured as a device for executing general-purpose entropy coding without limiting the signal processing contents of the modeling unit. For example, JP
In the case of EG, MPEG, etc., a portion corresponding to a modeling unit such as discrete cosine transform can be executed outside the LSI chip, and the LSI chip can execute entropy coding processing. Further, since various encoding targets are not limited, it is suitable for so-called multimedia signal processing, and can be used as a common entropy encoding device for images, voices, texts, line graphics, and the like.

(2) By using both the modeling unit and the entropy coding unit as LSI chips, a dedicated device limited to a specific coding target can be constructed. For example, JPEG, MPEG
In such cases, both the modeling unit such as the discrete cosine transform and the entropy coding unit can be configured as an LSI chip. The advantage of this case is that the signal transmission between
Wiring on the chip is possible, miniaturization of the device can be realized by mounting a buffering memory on the chip, and the compression rate can be improved by referring to the signal processing result of the modeling unit.

Next, a case where the encoding device 203 of the present invention is applied to a digital archive device will be described.

For example, the image data input for a work of art, etc. can be stored in high quality without deterioration,
It can be used for viewing reproduced images and for academic research. A digital archive device stores data for such a purpose. However, the input image signal is generally a multi-valued signal composed of a plurality of color signals, and it is difficult to efficiently compress it without degrading it.

By using the encoding device 203 of the present invention, a high compression rate can be realized with the configuration shown in FIG.

As the structure of the modeling unit, for example, a signal predicted by referring to a signal accumulated in a buffer memory,
There is a method of calculating a difference value from an actually input signal.
Then, the occurrence probability of the difference value can be predicted, and the entropy coding of the difference value can be performed. Generally, the difference value is 0
Since the amplitude is +/- around the center, the shortest code is assigned to the difference value 0, and the long code is assigned in the order of the amplitude of +-. Here, in the present invention, as shown in FIG. 30, it is data in the middle of the signal processing procedure of the modeling unit. (1) Statistical properties of the signal to be coded (2) Appearance distribution of difference values The occurrence probability can be obtained by measuring the characteristics and the like, and the code can be generated. Furthermore, the compression rate can be improved by expanding the information source that combines the signals of a plurality of pixels. If the encoding apparatus 203 of the present invention is used, when it is determined that signals having a difference value of 0 occur successively, it is possible to set a procedure such that a plurality of difference values are combined and converted into one code. Easy to implement. In an image having no noise component such as computer graphics, the difference value due to prediction is often 0. In such an image signal, the compression rate can be improved by combining the occurrence probabilities of a plurality of image signals to expand the information source.

As an example of a specific procedure based on the above, (a) as initialization, the reference signal is initialized.

(B) The prediction probability value is set from the reference signal, and the first signal level is used as the prediction signal.

(C) The difference value between the input signal and the predicted signal is calculated.

(D) Convert the difference value into a code.

(E) Repeat the procedure from step (b) for the next signal. If the signal input is completed, the end procedure is executed and the process ends.

As a storage device, a CD-ROM, a DVD,
A hard disk or a file device connected to a communication network can be used. In such a digital archive device, an encoding device and a decoding device are constructed separately, but the code generation procedure needs to have consistent processing contents. Therefore, the procedure (program) content for decoding is stored together with the compressed data, and the playback device can play back the compressed data by using the content for decoding. Next, application to run length coding will be described.

For example, in the case of raster scanning a black-and-white binary image such as a facsimile, the signal level of each pixel is either 0 or 1, and the image data is obtained by the continuous number (run length) of signals of the same level. Can be expressed. The occurrence probabilities of 0 and 1 can be predicted by referring to the adjacent pixel signals that have already undergone the encoding process. The present invention can be applied to such an encoding target as follows.

In the MH (Modified Huffman) method which has been conventionally used as a basic encoding method for facsimiles, a code table itself for replacing run lengths is prepared in advance as a standard. For this reason, there is an advantage that the transmission side device and the reception side device can commonly use it for encoding all kinds of images. However, on the other hand, there is a demerit that the compression performance is not sufficiently obtained when the image property is different from that when the code table is created. For example, when the run length is converted into a code while performing a raster scan of an image, the number of pixels in one line is constant, so the run length distribution changes depending on the base point of the run, but the fixed code table is used. The compression ratio may be deteriorated due to the use of. Further, since the character image and the dither image use the fixed code table even though the run length occurrence distributions are different, the compression ratio may be significantly inferior in the latter case.

On the other hand, if the coding / decoding apparatus of the present invention is used, code generation can be performed while measuring the run length property, and a high compression rate can be achieved.

As shown in FIG. 14, the probability setting unit measures the frequency of appearance of run lengths by referring to the signals for which the encoding process has already been completed. This signal processing can be commonly executed by both the encoding device and the decoding device. By the way, in the initial state where the encoding process is started, the occurrence distribution of the event (run length) is unknown. Further, when the number of signals to be referred to is small, the occurrence distribution for all the occurring events (run length) may not be measured. In these cases, the following means can be used.

(A) At initialization, some initial value is preset in the measuring means.

(B) The probability value of the entire event is obtained from the number of measurements that have already been measured.

(C) The occurrence probability is calculated for every fixed number of measurements.

(D) Apply some function or pattern to set the probability.

FIG. 31 illustrates a case where the frequencies for all run lengths cannot be measured because the number of measurements is small. Therefore, a probability distribution is set by, for example, a normal distribution function so as to complement the measurement result, and a certain probability value is replaced for a run length or more. The merits of setting such probability values are summarized below. (1) Since the probability distribution is expressed by the normal distribution function, the ranking of the probability values can be calculated from the function, and the operation for calculating the code length is a logarithmic operation of the normal distribution function and has a simple form.

(2) Since the occurrence of many actual events can be approximated by a normal distribution, the complementation error is small.

(3) It is possible to assign the same predicted probability value to a certain rank or lower of the ranked events while keeping the principle that the cumulative probability value does not exceed 1.0. For these events, a code length can be constantly assigned based on the code generation procedure described above. As a result, the distinction between these codes can be determined by the difference in the numerical size.

Next, the improvement of the compression rate of the JPEG (MPEG) system will be described.

The present invention can also realize the effect of improving the compression rate by replacing the conversion unit for the Huffman code used in the existing standard encoding method. For example, in an image coding method known as JPEG or MPEG method, a DCT (discrete cosine transform) section, a quantizing section, a scanning section, etc. are used in a modeling section with a block composed of a plurality of pixels as an encoding unit. It is composed of a signal processing unit. Then, the entropy coding unit converts the run length and the signal amplitude into Huffman code according to the scan order. Here, the efficiency can be improved by replacing the Huffman code conversion unit at the final stage with the adaptive code conversion unit according to the present invention.

FIG. 32 shows a basic configuration of a processing device that executes the JPEG system. The signal target to be converted to the code is
It is a series of 64 signals obtained by subjecting 8 × 8 two-dimensionally arranged signals, which are the signal processing units of the modeling unit, to DCT and quantization. DCT is a kind of two-dimensional frequency conversion, and quantization is a process of weighting frequency components. In the conventional method, these signal sequences have a signal value of 0.
Is combined with the run length indicating the number of continuous signals and the level of the signal (other than 0) at the run delimiter is converted into the Huffman code. Here, there is a case where encoding by a single code table is performed as a factor that reduces the compression rate in the conventional Huffman code conversion unit. For this reason, (a) the characteristics of the target image signal, the characteristics of the frequency component, etc. are not taken into consideration. (B) It is not taken into consideration that the characteristics of the run length and the signal level differ depending on the start position of the run in the signal sequence.

(C) A Huffman code is assigned to an event that combines different properties such as run length and signal level.

Even if the above problem is solved by using the Huffman code, there are the following problems.

(A) A plurality of code tables are prepared in advance, and the code tables are sequentially switched while referring to the signal series of adjacent blocks. However, the code table becomes huge and the device scale increases.

(B) The code table is sequentially generated with reference to the signal sequence of the adjacent block. However, the processing time becomes long and the device scale increases.

FIG. 32 shows a device configuration for improving the compression ratio of the JPEG system. Assuming that the same modeling unit as in the JPEG system is used, a case will be described in which the run length of continuous 0 and the signal level at the break point of the run are converted into a code according to the scan order. In order to speed up the signal processing, it is effective to dispose a buffer memory for accumulating at least one block of signals before and after the DCT unit and the quantization unit, and perform the operations of the respective units in parallel. Various signals stored here are referred to in order to improve coding efficiency,
The following signal processing can be performed.

(A) As a referenceable signal sequence, a signal of the object to be encoded by prescan, a signal which has already been encoded, a different color signal of the same image in the case of a color image, and a moving image in the case of a moving image There are different frame signals, etc. By correlating the already processed signal with the run length distribution and the signal level distribution, it is possible to predict the run length distribution and the signal level distribution to be encoded using the reference signal.

(B) The run length and the occurrence probability of the signal level can be changed depending on where the run base point is located in the 8 × 8 signal sequence. This is because the maximum length of the signal sequence is limited to 64 (= 8 × 8), and since the encoding process proceeds according to the scan order, the run length occurrence probability varies depending on the run base point. This is because it is natural. If the same code table is always used regardless of the origin of the run, the probability of occurrence of the run length is not effectively used, and the compression efficiency is reduced.

(C) For both the run length and the signal amplitude, independent occurrence probabilities are predicted to generate a code.

In any case, since there is no signal that can be referred to at the initial point, initialization processing is performed, such as using some set value or storing and using information related to a conventionally processed image. I will decide. By referring to the signals as described above, the timing for performing adaptive code generation is as follows: (a) Code generation is performed at each code conversion timing of the run length and the signal level.

(B) Code generation is performed in units of one or a plurality of blocks.

(C) Code generation is performed in units of one screen (color signal in the case of a color image).

(D) In the case of targeting an image composed of a plurality of temporally consecutive image frames, some frames are similar because the temporally adjacent image frames have similar properties. A code is generated for each.

An example of a specific processing procedure will be described below.

(A) At the initial point of the encoding process of the 8 × 8 signal sequence, the base point of the run is set to the signal (DC component) having the lowest frequency component. The run base point position RS = 0 is set.

(B) The number j of which the signal component is 0 is measured from the run base point RS in the scanning order. When the signal level SS which is not 0 is detected, the run length RL = j is set and the process proceeds to the next step. However, if the run reaches the end during the measurement, proceed to (d).

(C) The run length distribution is predicted based on the run base point position RS, and the code generation procedure is executed until the run length RL code is generated. Subsequently, the signal level distribution is predicted based on the run base point RS, and the code generation procedure is executed until the code of the signal level SS is generated. As described above, the signals other than the run base point can be referred to for the prediction process.

(D) When the run reaches the end of the signal sequence,
Termination processing (for example, addition of a termination code) is performed, and the process ends. If it is not the end point, the new run base point is set to RS + RL and the process returns to (b).

The code generated when the above adaptive code conversion unit is used is not data compatible with the standard method, but since both are block codes, it is 1: 1 by some means. The code conversion of can be performed.

Next, application to dictionary type encoding will be described.

An encoding method in which a character string having a high appearance probability is sequentially registered and the order of registration is represented as a code is targeted for a text or the like formed by a sequence of character codes.
It is known as dictionary type encoding (for example, LZ encoding). The configuration of the registration means and the signal processing procedure can be classified into the modeling unit. On the other hand, in the part corresponding to the entropy coding unit, a method of directly using the rank value of registration or a method of converting the rank into a Huffman code has been proposed as a code. From the viewpoint of compression rate, it is a fact that the uniqueness of the occurrence probability between characters is already known, and the statistical value is created sequentially based on the input character code, and the probability distribution is created based on this content. By doing so, efficient encoding processing can be performed.

Here, if the conventional Huffman code is applied, there is a problem that if the code table is fixed, the adaptability is lacking, and if the codes are sequentially generated, the processing load increases.

On the other hand, when the compression / expansion device of the present invention is used, code generation can be performed by utilizing the following features of dictionary type coding.

(A) Encoding a character string in which a plurality of characters are combined is nothing but expansion of the information source.

(B) Sequential registration of character strings is nothing but measuring the occurrence probability.

By utilizing the above property, a link to which transition probabilities of character strings are assigned as shown in FIG. 33 is created,
An efficient block code can be generated by the following procedure.

(A) A candidate character string is prepared as a data structure connected by a link having a transition probability between characters as an attribute.

(B) If the input character string matches the candidate character string, a code based on the transition probability of the candidate character string is generated.

(C) The transition probability can be sequentially updated by the character string input as the actual encoding target. Here, the transition probability updating process can be performed for each character or after a certain number of characters are input, and the like, and the configuration of the algorithm and means can be adopted, but the present invention is not particularly limited thereto. .

The judgment of the delimiter of the input character string is performed by a space,
It can be determined by a character indicating a delimiter such as parentheses, and can be determined as a delimiter when the transition probability between characters is lower than a threshold value. The candidate string is
By deleting links with low transition probability from time to time, it is possible to reduce the scale and complexity. The configuration of the algorithm and means for deleting this link is not particularly limited. Transition probabilities below a certain level can be replaced with uniform transition probabilities.

(D) The sign bit setting can be calculated by sequentially multiplying the transition probabilities between characters, but the load for the calculation may increase. However, since the ratio of the number of all candidate character strings to the number of appearance of the corresponding character string is the transition probability, it is sufficient to perform one division using both numbers in order to obtain this ratio. . Furthermore, if the transition probability is updated at the timing when the number of input character strings becomes 2 N (for example, 1024), the division can be replaced with a binary N-bit shift. Based on the probability value obtained in this way,
Code generation can be performed at high speed and a high compression rate can be realized.

Next, the relation with the arithmetic code will be described.

By expanding the information source using the code generation algorithm according to the present invention, the generated code can be made to have a property close to an arithmetic code. As a result, it is possible to realize a compression rate as close as possible to the entropy.

For example, for all the signal sequences to be coded, one probability value is calculated by combining the occurrence probabilities of the events that actually occurred, and the coding process for this probability value is executed. You can

It should be noted here that, as described above, the code according to the present invention can be uniquely distinguished as the magnitude of a numerical value. This is 0.0, as shown in the explanation of arithmetic codes.
As described above, it means that the code can be generated by dividing the area on the number line in the range smaller than 1.0 by the probability value of the event and indicating the position of the upper end or the lower end of the divided area as a numerical value. As described in the code generation procedure, when the code is generated in the order of the magnitude of the probability value, the magnitude as the numerical value of the generated code is such that the number linear shape changes from 0.0 to 1.0. You will go up in sequence. FIG. 34 shows division of the number line. Performing the encoding process for each individual signal means that the division of the number line is independently performed for each code as shown in FIG. On the other hand, expanding the information source by combining a plurality of signals corresponds to repeatedly performing division on the number line as shown in FIG. 34B with the occurrence probability of the signal. The principle of this division is the same as the arithmetic code and the generation procedure, but is characterized by the following points.

(A) The number of signals to be encoded can be arbitrarily set to give the code sequence the characteristics of the block code. That is, the number of bits of the generated code can be set clearly.
This is not possible with arithmetic codes.

(B) As an application example of the arithmetic code, ITU
(International Telecommunication Union) is T.82
There is a JBIG encoding system recommended as. This is a coding method determined based on the idea of arithmetic codes for the purpose of realizing higher efficiency than conventional MH codes for black-and-white binary images. However, the signal processing procedure has been greatly simplified from the viewpoint of practicality. For example, for the purpose of speeding up, the occurrence probability of each event is replaced with a power of (1/2), and then the occurrence probability is calculated. The above example is premised on a binary signal as an encoding target, but in the case of a multilevel signal, the complexity is increased, and therefore further simplification is required. As described above, the theoretical compression efficiency of the arithmetic code is high, but from a practical point of view, a reduction in efficiency due to simplification is inevitable.

On the other hand, in the present invention, the occurrence probability of an event is correctly used, and it can be effectively used even for a multilevel signal,
Further, the effect of improving the compression rate by expanding the information source can be realized. The number of signals to be combined can be set in cooperation with the operation of the modeling unit, such as one scan line of an image or a block composed of a plurality of pixels.

Next, the N-ary number encoding will be described.

Although the case of the binary code has been described so far, it is needless to say that the same effect can be obtained by generating the N-ary code by using the present invention. Examples of such an apparatus include a memory that can store N values per bit, and an N-value communication path that performs communication by amplitude / phase modulation (modem). Conventionally, a multi-level signal is generated by combining a binary code with a plurality of bits, but in the present invention, an N-ary code can be directly generated. For example, when data transmission is performed at three voltage levels of +1, -1, 0, efficient data transmission can be realized by using a ternary code. Furthermore, with the encoding process, the appearance of each of the three types of signals becomes almost random, so that the voltage can be balanced to 0 on average. In addition, it is easy to adaptively change the value of N according to the characteristics (noise, LR characteristics, distance, etc.) on the transmission path, and highly reliable data transmission can be performed by combining with various error correction methods. Can be realized. For example,
When a transmission error occurs, the resistance to the error can be increased by retransmitting the generated code with a smaller N value than the previous time.

The signal representation of the transmission line or the temporary storage buffer until the generated N-ary code is passed to the N-value memory or the N-value communication line is, for example, a binary number and may have a value different from N. .

In this case, the N-ary code is temporarily set to 2
It can be solved by expressing it in a decimal number. Thus, signal processing can be performed in combination with code conversion known as channel coding.

[0169]

As described above, according to the present invention, it is possible to generate a code having an optimal and instantaneously decodable property at high speed. Therefore, it is possible to sequentially generate a code adapted to the signal property and realize a high compression rate. Further, the expansion of the information source makes it possible to easily improve the compression rate.

[Brief description of drawings]

FIG. 1 is a diagram showing a basic configuration of an encoding device.

FIG. 2 is a diagram showing a configuration of an encoding system.

FIG. 3 is a diagram for explaining an outline of code generation processing.

FIG. 4 is a diagram showing memory storage for code generation.

FIG. 5 is a diagram showing an overall process of code generation.

FIG. 6 is a diagram showing a process of initial setting.

FIG. 7 is a diagram showing a code generation process.

FIG. 8 is a diagram showing a process of encoding error calculation.

FIG. 9 is a diagram showing the contents stored in a memory.

FIG. 10 is a diagram showing a relationship between generated codes.

FIG. 11 is a diagram for explaining an outline when a probability function is used.

FIG. 12 is a diagram showing a process of initial setting when a probability function is used.

FIG. 13 is a diagram showing a process of code generation when a probability function is used.

FIG. 14 is a diagram for explaining the configuration of image data.

FIG. 15 is a diagram for explaining initial setting of probability when image data is encoded.

FIG. 16 is a diagram showing an entire code generation process when obtaining a sequential probability.

FIG. 17 is a diagram showing a code generation process when obtaining a sequential probability.

FIG. 18 is a diagram showing a configuration of a decoding system.

FIG. 19 is a diagram showing a configuration of an encoding device.

FIG. 20 is a diagram for explaining the outline of the decoding process.

FIG. 21 is a diagram showing the entire process of decoding.

FIG. 22 is a diagram showing a decryption process.

FIG. 23 is a diagram showing a decoding process when a probability function is used.

FIG. 24 is a diagram showing an entire decoding process when obtaining a sequential probability.

FIG. 25 is a diagram showing a code generation process for obtaining a sequential probability.

FIG. 26 is a diagram showing the configuration of an encoding / decoding device.

FIG. 27 is a diagram showing a configuration of an embedded LSI.

FIG. 28 is a diagram showing a configuration of a system using virtual memory.

FIG. 29 is a diagram showing a configuration of a digital archive device.

FIG. 30 is a diagram showing an application example to run length encoding.

FIG. 31 is a diagram showing an example of application of the JPEG (MPEG) method to improve the compression rate.

FIG. 32 is a diagram showing an application example to dictionary type encoding.

FIG. 33 is a diagram showing a relationship with arithmetic codes.

─────────────────────────────────────────────────── ─── Continuation of front page (58) Fields surveyed (Int.Cl. ⁷ , DB name) H03M 7/40 H04N 1/41 H04N 7/24

Claims (4)

(57) [Claims] 1. A coding method for generating a sign of the input signal, the codeword based on the occurrence probability of a defined event to occur in advance
And calculate the cumulative probability based on the code length of the generated codeword.
Therefore, from the code length of the generated codeword,
The error is calculated, and the probability value based on the calculated coding error is calculated for other events.
An encoding method characterized by distributing to occurrence probabilities and generating a code from the code length and the cumulative probability . 2. Decoding for recovering a signal from an input code
In the method, the code word is based on the probability of occurrence of a predetermined event.
And calculate the cumulative probability based on the code length of the generated codeword.
Therefore, from the code length of the generated codeword,
The error is calculated, and the probability value based on the calculated coding error is calculated for other events.
Distribution to occurrence probabilities, generating a code from the code length and the cumulative probability, and associating an input code with the generated code.
Characterizing decoding method . 3. A coding device for generating a code of an input signal.
, A codeword based on the occurrence probability that a predetermined event will occur.
Based on the code length of the codeword generated by the code generation unit and the code generation unit
The cumulative probability is calculated, and the code is calculated from the code length of the generated code word.
The coding error associated with the coding is calculated, and the calculated coding error is calculated.
A code that divides the probability value based on the difference into the occurrence probabilities of other events.
A coded error calculation unit, and the code generation unit calculates a code from the code length and the cumulative probability.
Encoding apparatus and generates a. 4. Decoding for recovering a signal from an input code
In the device, a code word based on the probability of occurrence of a predetermined event
Based on the code length of the code word generated by the decoding unit and the decoding unit that generates
The product probability is calculated, and the code is calculated from the code length of the generated code word.
Encoding error due to encoding, and the obtained encoding error
Coding for Distributing Probability Values Based on Attribute to Occurrence Probabilities of Other Events
An error calculation unit, and the decoding unit calculates a code from the code length and the cumulative probability.
Generate and associate the input code with the generated code
Decoding device, characterized in that that.