JP2827982B2 - Data compression control method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] 1. Field of the Invention [0002] The present invention relates to a method for controlling data compression processing.

[0002]

2. Description of the Related Art Data compression processing is effective for efficiently storing and communicating data. Data compression is achieved by exploiting the statistical properties of the data. A data structure that holds the statistical properties of this data is called a “model”. In the adaptive data compression method, this model is dynamically updated as the compression progresses. As the compression process progresses, compression can be performed using a model that better captures the properties of data, so that the adaptive data compression method can obtain an excellent compression ratio as the compression proceeds. The universal compression method, which is an effective compression method for many types of data, is often an adaptive data compression method because it is necessary to dynamically grasp the nature of the data. As a distortion-free universal compression method capable of completely restoring original data, a Lempel-Jib method, a method combining a context tree and an arithmetic code, and the like are known. These schemes are also adaptive data compression schemes that dynamically update the model.

In the case of the Lempel-Jib method, the model is in the form of a dictionary or a buffer. The method using a dictionary is LZ78 type, and the method using a buffer is LZ7.
It is called type 7.

The dictionary in the LZ78 type has a structure in which a character string in input data is cut out according to a certain rule, and an index is assigned. If the string in the dictionary reappears, compression is achieved by replacing the string with an index in the dictionary.

The LZ77 type buffer has a structure for storing the latest input character string which has already been compressed. If the string in the buffer reappears, it is replaced by an index consisting of two pairs: the position where the string starts (how many characters before the encoding position?) And the length. Compression is achieved. LZ77
Basically, a type can be processed if there is a buffer like this. However, in order to speed up the compression process, a data structure for speeding up a character string search in the buffer is required, and the data structure is sometimes called a dictionary.

[0006] For details of the Lempel Jib method, see the book text compression (Tex
t Compression, 1990, Prent
ice Hall Inc., pp. 206-243). The initial state of the dictionary and buffer is basically empty, and compression is started. As the compression proceeds, a new character string is registered in the dictionary, and a dictionary that captures the characteristics of data is generated.

On the other hand, in the method using a context tree, the model is in the form of a context tree and the appearance frequency of each character under each context. Actual encoding is performed by arithmetically encoding each character based on the frequency of appearance. A context tree is a set of contexts. Each time a new context appears, a new context is registered in the context tree as needed. When each context appears in the input data,
The frequency of occurrence of the character in that context is updated according to the next character.

The method using the context tree is described in detail in the above-mentioned book text compression (pp. 140-166). The context tree is basically empty and starts to be compressed.As the compression progresses, a new context tree is added to the context tree, the frequency of occurrence of characters under each context is updated, and the nature of the data is well understood. A context tree is generated.

In the above-described adaptive data compression method, a good model is constructed as the data compression progresses, so that the processing proceeds to some extent from the time when the data compression is started. Enables more effective compression.

At the time of restoring data compressed by the adaptive data compression method, the restoration cannot be performed correctly unless the same model as that at the time of compression is used. For this purpose, it is necessary to set the initial state of the model equal to that at the time of compression, and to update the model as in the case of compression.

[0011]

The adaptive data compression method is a compression method capable of obtaining an excellent compression ratio. However, if an error such as bit inversion occurs in the compressed data, the subsequent data cannot be restored at all. There is a problem.
If the compression codeword is of variable length, errors may occur at the boundaries of the compression codeword due to bit inversion, in which case the data may be catastrophically damaged. In addition, even if only one compression codeword is decompressed while it is incorrect, the corresponding decompressed data will be different.In the case of the adaptive data compression method, the model is updated based on the decompressed data that has been decompressed incorrectly. Therefore, the error may propagate without stopping. By combining error correction coding, the occurrence of errors can be suppressed to some extent, but even if a large number of errors occur, correction becomes impossible.

To solve such a problem, conventionally, as shown in FIG. 2, a method of dividing data into small blocks and independently compressing each block has been adopted. Even if restoration of one block fails, the remaining blocks can be restored correctly. By returning the model to a predetermined initial state at the start of compression of each block, each block can be compressed independently. JP 05-252047 A, JP 05-252048 A
The method disclosed in the publication is essentially the same type of method. When applied to a communication system, the restoration side only needs to request retransmission of only the block in which an error has been detected, and there is no need to restore the entire data from the beginning.

However, in the case of the adaptive data compression method, compression is performed while constructing a model, so that an excellent compression ratio is achieved for large data for which the model can be sufficiently grown for the first time. In the case where compression is performed independently in small block units, a predetermined initial state is set at the head of the block, so that a model suitable for the data cannot be constructed and a sufficient compression ratio cannot be obtained. There was a point. On the other hand, when a block is enlarged, there is a problem that the loss data when the block cannot be restored increases.

[0014]

Means for Solving the Problems (1) The present invention relates to an adaptive data compression system that uses a model that is a data structure representing statistical properties of data and performs compression while dynamically updating the model. In the data compression control method, input data is divided into several blocks, compression is performed in units of the blocks, clusters in which a certain number of adjacent blocks are put together are formed, and different clusters are independently compressed, and When the head of the cluster is compressed, the initial state of the model is set to a predetermined setting. When the block other than the head of the cluster is compressed, the initial state of the model is the block. The setting is determined using the first block of the cluster to which.

(2) The size of the block is constant.

(3) The size of the block is variable, and the size of the block is controlled so that the size when the block is compressed becomes equal.

(4) The present invention uses a buffer for storing input data that has already been compressed as the model,
When the character string in the buffer appears again in the input data, the start position of the character string, a set of indices indicating the length is set as a codeword for the character string, and the buffer is always compressed. When updating so that the latest data is stored and compressing the head block of the cluster, the initial state of the buffer is set to a predetermined setting, and the block other than the head of the cluster is compressed. The initial state of the buffer is the state of the buffer when the compression of the first block of the cluster to which the block belongs has been completed.

(5) The present invention uses a dictionary in which an index is made to correspond to a character string in input data which has already been compressed as the model, and when the character string in the dictionary appears again, The index of the character string in the dictionary is a code word for the character string, the dictionary is updated by assigning and registering an index to a character string newly appearing in the input data, and updating the block at the head of the cluster. When compressing, the initial state of the dictionary is set to a predetermined setting, and when compressing the block other than the head of the cluster, the initial state of the dictionary is set at the head of the cluster to which the block belongs. The dictionary is in a state when the compression of the block is completed.

(6) The present invention uses a structure in which a set of contexts, which are character strings appearing in input data as the model, and an appearance frequency of each character in the input data under the context correspond to each other. Encoding each character in the input data based on the frequency of occurrence under the context, and updating the set of contexts by adding a character string newly appearing in the input data to the set of contexts And updating the occurrence frequency each time a character appears under the context in the input data, and compressing the block at the head of the cluster. The initial state of the appearance frequency of the character is set in a predetermined manner, and when the block other than the head of the cluster is compressed, it is obtained when the compression of the head of the cluster to which the block belongs is completed. The context Characterized in that the initial state of each appearance frequency of under set and context.

[0020]

When the present invention is used, compression is performed independently for each cluster, which is a set of a fixed number of blocks. Therefore, even if restoration of one cluster fails, the remaining clusters can be restored correctly. . In addition, if the restoration of the first block in the cluster fails, the rest of the blocks in the cluster cannot be restored. Can be restored correctly. When compressing a block other than the head of a cluster, compression is performed after a model is constructed to some extent using the head block, so that a compression ratio superior to the size of the block can be obtained.

[0021]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention is applied to an adaptive data compression system that uses a model, which is a data structure representing the statistical properties of data, and performs compression while dynamically updating the model. The model is updated from an initial state as the compression progresses. The model is realized in the form of, for example, a dictionary or an appearance frequency table.

A major feature of the present invention is that compression is basically performed on a block basis, but a cluster in which a certain number of adjacent blocks are combined is formed. Let K be the number of blocks in a cluster, and assign 0 to (K
Numbers up to -1) are assigned, and are referred to as block numbers.

An embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a flowchart showing the flow of the process of the present invention.

First, it is determined whether the block to be compressed is the head of a cluster (step S1). This can be determined based on whether the block number of the block to be compressed is 0 or not. The block number at the start of compression is 0.

If it is the first block of the cluster (block number = 0), the initial state of the model is set to a predetermined state (step S2). When the appearance frequency table of each character is used as a model, for example, the appearance frequency of each character is set to 1
Set to.

If it is not the first block of the cluster,
The model is set in a state where the compression of the head block of the cluster to which the block belongs has been completed (step S3). When the appearance frequency table of each character is used as a model, the appearance frequency table is updated as the compression of the first block progresses, and compression of the block is started with the appearance frequency table at the end of the compression of the first block as an initial state. There are two methods for this initial state setting process. One method is to store the current model in another area when the compression of the first block of the cluster is completed, and to compress the stored model when compressing the remaining blocks in the cluster. This is a method of copying to the area of the model to be used. Another method is a method of performing compression processing of the first block of a cluster and compressing the model before compressing each block, and starting compression of each block when the state of the first block compression is completed. However, in the process of the head block compression process, only the model is updated, and no codeword is actually output. As compared with the conventional method shown in FIG. 2, the former method requires extra memory, and the latter method requires twice the compression time. The former method seems to be effective at present when the memory capacity is becoming large and inexpensive.

When the initial state of the model is determined, each block is compressed (step S4). In the process of compression, the model is updated sequentially.

Before moving to the next block, it is checked whether data to be compressed still exists (step S).
5). If there is, the block number is incremented by 1 (if it becomes K, it is set to 0), and the process shifts to compression of the next block (step S6). If not, the compression ends.

There are two ways to determine the size of a block. One is a method of making each block before compression the same size, and the other is a method of making each block after compression the same size. Since the cluster is composed of a fixed number of blocks, the former necessarily has the same size as the cluster before compression, and the latter has the same size as the cluster after compression. When actually writing the compressed data on the medium, error correction coding or error detection coding using Reed-Solomon code or the like is performed. Error correction codes such as Reed-Solomon codes are described in the book "Code Theory" (1990,
IEICE, pp. 151-187). In error correction coding using a Reed-Solomon code or the like, coding is performed in units of blocks, but the size of a block serving as a unit of error correction coding and a block in compression may be irrelevant.

When the size of the block before compression is fixed, the compression processing is easy and is suitable for the parallel compression processing in cluster units. Only the last block of input data may be different in size than the other blocks.

When the size of the block after compression is fixed, when the size of the compressed data reaches the size, the size of the block in the original data is determined, and the process shifts to compression of the next block. Only the last compressed block may be different in size from other compressed blocks. This method is
In addition to being suitable for parallel restoration processing in cluster units, the compressed block has a fixed length, so that even if an error occurs in a block, there is a great advantage that it can always be shifted to the next block or cluster. FIG. 3 shows an example of the format of original data and compressed data when the size of the compressed block is constant. In FIG. 3, the cluster is composed of four blocks.
When the first block (block 0) having the block number 0 is compressed, the initial state of the model is set to a predetermined setting, and compression is performed. When the size of the compressed data reaches a predetermined size, the block is divided there, and the subsequent data is the next block. At the time of compressing the blocks (block1, block2, block3) of block numbers 2, 3, and 4, the block (b) of block number 0 of each cluster
The compression is started with the state of the model at the end of the compression of (lock 0) as an initial state.

Next, the decompression processing of the data compressed by the present invention will be described. FIG. 4 is a diagram showing the flow of this processing. The number of blocks in the cluster is K.

Before decompressing each compressed block, it is determined whether an error is mixed in the compressed block (step T1). This can be performed using an error correction coding method and an error detection coding method. These techniques are described in detail in the aforementioned book "Code Theory".

If it is determined that an uncorrectable error is mixed in the compressed block, it is first determined whether or not the block is the head block of the cluster (step T2). This can also be determined based on whether the block number is 0, as in the case of compression. The block number at the start of restoration is set to 0.

If an error is found in the first block of the cluster, the information is notified (step T).
3), restoration of the cluster is interrupted, and the process shifts to restoration of the next cluster (step T5). The block number is 0. If it is determined that an error has occurred in a block other than the head of the cluster, the information is notified (step T4).
The process proceeds to the restoration of the next block (step T6). The block number is incremented by 1 (if the block number = K at this time, the block number is set to 0).
When the size of the compressed block is fixed, it is always possible to move to the restoration of the next block or cluster.

Even when it is determined that no error is mixed in the block, first, it is determined whether or not the block is the head block of the cluster (step T7). If the block is the first block of the cluster, the model is set to a predetermined initial state (step T8). When using the appearance frequency table of each character as a model, for example, all are set to 1. If the block is not the head block of the cluster, the model is set to the state of the model when the head block of the cluster is restored (step T).
9). When using the appearance frequency table of each character as a model,
As the restoration of the head block progresses, the appearance frequency table is also updated, and restoration of the block is started with the appearance frequency table at the end of restoration of the head block as an initial state. In this processing, two methods are conceivable as in the case of compression. There are a method of retaining the state of the model at the end of restoring the first block in another area, and a method of restoring the first block again (recovery data is not output). When the initial state of the model is determined, the compressed block is restored (step T10), and the process proceeds to the next block (step T11). The block number is incremented by 1 (if the block number = K at this time, the block number is set to 0).

If there are no remaining compressed blocks, the decompression process ends (step T12).

Next, an example of a combination of the data compression control method of the present invention and a specific compression method will be described.

A case where the LZ77 type is applied as a compression method to the present invention will be described. In the LZ77 type, a buffer is used as a model.

First, the LZ77 type will be described. The input data is x (0) x (1) x (t)... X (n). At this time, compression by the LZ77 type is performed as follows. (1) Initialize the buffer. j = 0 (2) The following processing is repeated: (i) Match between x (j) x (j + 1)... and the character string in the buffer to obtain the longest matching string.

(Ii) If the length m of the longest matching string is smaller than the threshold, the first character x (j) is output as it is. m = 1.

If the length m of the longest matching sequence is larger than the threshold value, the starting position of the longest matching sequence (how many characters before the encoding point?) And an index representing the length m are x (j) x
(J + 1)... X (j + m-1) are output as codewords.

(Iii) A character string x (j)... X (j + m−) obtained by shifting the character string in the buffer by m characters and completing the encoding.
Insert 1) into the buffer.

(Iv) j ← j + m.

(V) End when j> n. By the process (3) of (2), the buffer is updated so as to store the latest input data after compression.
In (ii) of (2), it is necessary to devise a method such as using a flag to determine whether the code word is the character itself or an index indicating the position and length.

The case where the present invention is applied to this compression method will be described.

First, it is determined whether or not the block to be compressed is the head of the cluster (step S1). This can be determined from the block number as in the case of the description of FIG.

When compressing the first block of the cluster, the initial state of the buffer is set to a predetermined state, for example, the buffer is set to an empty state as shown in FIG. 5A (step S2). When compressing a block that is not the head of a cluster, the buffer is set so that the compression of the head block of the cluster is completed (step S3). For this purpose, it is necessary to store the state of the buffer at the end of the first block compression in another area. The first block is x
(S) x (s + 1)... X (t), and if the buffer is sufficiently large, the head block is directly stored in the buffer as shown in FIG. 5B. The means for searching for a character string in the buffer also needs to be set to the state when the compression of the first block is completed.

After the setting of the buffer is completed, compression is started (step S4).

When the compression of the block is completed, it is determined whether or not the remaining data exists (step S5). If there is remaining data, the block number is incremented by 1 (step S6), and the process proceeds to compression of the next block.

The case where the LZ78 type is applied as a compression method to the present invention will be described. The LZ78 type uses a dictionary as a model.

First, the LZ78 type will be described. The input data is x (0) x (1) x (t)... X (n). At this time, compression by the LZ78 type is performed as follows. (1) Initialize the dictionary. j = 0 (2) The following process is repeated: (i) Match between x (j) x (j + 1)... and the character string in the dictionary to find the longest matching string.

(Ii) The longest matching sequence (x (j) x (j +
1)... X (m)) in the dictionary are output as codewords.

(Iii) x (j) x (j + 1)... X
Allocate a new index to (m) x (m + 1),
Register in the dictionary.

(Iv) j ← m + 1.

(V) End when j> n. The process (3) of (2) means that the dictionary is updated by newly registering a character string obtained by extending the character string registered in the dictionary by one character into the dictionary. .

The case where the present invention is applied to this compression method will be described. Let the set of characters be {a, b, c, d}.

First, it is determined whether or not the block to be compressed is the head of the cluster (step S1). This can be determined from the block number as in the case of the description of FIG.

When compressing the head block of a cluster, the initial state of the dictionary is a predetermined state, for example, as shown in FIG.
It is assumed that all the characters are registered in the dictionary as shown in (a) (step S2). When compressing a block that is not the head of a cluster, a dictionary is set when the compression of the head block of the cluster is completed (step S3).
For this reason, it is necessary to store the state of the dictionary at the end of the first block compression in another area. The first block is a
If the character string is bbcbbbcbac,
The state shown in (b) is the state of the dictionary when this block is compressed. For blocks other than the head, compression is started after setting this dictionary to this state.

After setting the dictionary, compression is started (step S4).

When the compression of the block is completed, it is determined whether or not the remaining data exists (step S5). If there is remaining data, the block number is incremented by 1 (step S6), and the process proceeds to compression of the next block.

A case where an adaptive data compression method is applied to the present invention using a data structure in which a context and a frequency of appearance of characters under the context are modeled. Although it is possible to dynamically update the context set, here, a model (first-order Markov model) in which the context set is fixed to a set consisting of one character will be described. This model has a structure in which an appearance frequency table is associated with each character, and the number of appearance frequencies is a set of counters for counting the number of appearances of each character under the character.

First, an adaptive data compression method using a first-order Markov model will be described. Now, it is assumed that the set of characters of the input data consists of four characters {a, b, c, d}.
For each of {a, b, c, d}, an appearance frequency table including a counter for counting the number of appearances of the character appearing next to the character is prepared. Let c (x, y) be the number of times the character y appears after the character x. For example, for a, c (a, a), c (a, b), c (a,
c) and an appearance frequency table composed of four counters storing the values of c (a, d) is associated.

The encoding of the input data x (0) x (1)... X (n) is performed in the following flow. (1) Set the counter to the initial state. j = 0 (2) Repeat the following processing: (i) c (x (j−1), a), c (x (j−1),
b), c (x (j-1), c), c (x (j-1),
arithmetically encode x (j) using d). Let the application probability of (x (j) be c (x (j-1), x (j)) / c (x (j
-1)). Here, c (x (j-1)) = c (x
(J-1), a) + c (x (j-1), b) + c (x
(J-1), c) + c (x (j-1), d). ) (Ii) c (x (j−1), x (j)) ← c (x (j−
1)), x (j)) + 1 (iii) j ← j + 1.

(Iv) End when j> n. Here, x (-1) is a predetermined character (for example, a). The appearance frequency table is updated by the processing of (2)-(ii). Regarding the arithmetic coding of (2)-(i), the above-mentioned book “Text compression” (pp. 10)
2-139).

A case where the method of the present invention is applied to this compression method will be described.

First, it is determined whether or not the block to be compressed is the head of the cluster (step S1). This can be determined from the block number as in the case of the description of FIG.

When compressing the first block of the cluster, the initial state of the appearance frequency of each character is set to a predetermined setting. For example, as shown in FIG.
(Step S2). When the context tree itself is dynamically updated, for example, the context tree is also set to an empty state. When compressing a block other than the head of the cluster, the initial state of the appearance frequency of each character is set to a state where the head block ends as shown in FIG. 7B (step S).
3). In FIG. 7B, c0 (x, y) is the value (c (x, y)) of the counter at the end of the compression of the first block. When the context tree itself is dynamically updated, the context tree at the time when the compression of the first block is completed is set as the initial state.

When the setting of the counter is completed, the compression is started (step S4). The value of the counter is updated as the compression proceeds.

When the compression of the block is completed, it is determined whether or not the remaining data exists (step S5). If there is remaining data, the block number is incremented by 1 (step S6), and the process proceeds to compression of the next block.

[0071]

The following two assumptions are made.

When an uncorrectable bit (error) appears in a compressed block, the entire block cannot be restored (operation to restore the compressed block).

Uncorrectable bits occur randomly,
The probability is small enough.

The size of the compressed block is fixed.

Under this assumption, the average value of the number of unrecoverable bits in the case of using the present invention is calculated by the conventional data compression control method for independently compressing data in block units, and determining the size of the compressed block. It is almost equal to the value when the size of the compressed block of the invention is doubled. That is, the error propagation control ability by the compression control of the present invention is almost the same as the error propagation control ability when the size of the compressed block is doubled by the conventional compression control in which compression is independently performed in block units.

As a compression method, LZ77 type (1 character = 1 byte, buffer size = 8192, maximum value of matching string = 3
2. An experiment was performed using pointer value isometric coding. In the present invention, the size of the compressed block is 1024 bytes, and the number of blocks in the cluster is 16. In the conventional data compression control method incorporating the LZ77 type and performing independent compression in block units, the size of the compressed block is doubled to 2048 bytes. Comparing the two compression ratios, the present invention was often superior by about 2 to 5 points. That is, by using the present invention, it is possible to obtain a compression ratio superior to that of the related art by about 2 to 5 points with the same error propagation control ability.

Further, an experiment was conducted using the LZ78 type (one character = 1 byte) as a compression method. The size of the dictionary used in the LZ78 type was sufficiently large. In the present invention, the size of the compressed block is 10 as in the case of the LZ77 type.
In the conventional data compression control method, the number is set to 24 bytes, which is twice as large as 2048 bytes. The number of blocks in the cluster according to the present invention is also 16 as in the case of the LZ77 type. When the compression ratios of the two were compared, the present invention was often superior by about 1 to 4 points. In other words, by using the present invention, it is possible to achieve one-
An excellent compression ratio of about 4 points can be obtained.

[Brief description of the drawings]

FIG. 1 is a diagram showing a flow of compression processing control of the present invention.

FIG. 2 is a diagram showing a flow of conventional compression processing control.

FIG. 3 is a diagram showing formats of original data and compressed data when a compressed block is made equal length in the present invention.

FIG. 4 is a diagram showing a flow of a restoration process of data compressed under control of the present invention.

FIG. 5 is a diagram showing an initial state of a buffer when an LZ77 type is used as a compression method in the present invention.

FIG. 6 is a diagram showing an initial state of a dictionary when an LZ78 type is used as a compression method in the present invention.

FIG. 7 is a diagram showing an initial state of a counter when a context model (first-order Markov model) is used as a compression method in the present invention.

[Explanation of symbols]

 None

Claims (6)

(57) [Claims] 1. A data compression control method in an adaptive data compression system for performing compression while dynamically updating a model by using a model which is a data structure representing a statistical property of data. Is divided into blocks, and compression is performed on a block-by-block basis, a cluster is formed by grouping a certain number of adjacent blocks, and different clusters are independently compressed, and the first block in the cluster is compressed. In
The initial state of the model is set to a predetermined setting, and when compressing the block other than the head of the cluster, the initial state of the model is determined using the block at the head of the cluster to which the block belongs. A data compression control method, characterized in that: 2. The data compression control method according to claim 1, wherein the size of said block is fixed. 3. The data compression method according to claim 1, wherein the size of the block is variable, and the size of the block is controlled so that the size when the block is compressed becomes equal. Control method. 4. A buffer for storing input data that has already been compressed is used as the model. If a character string in the buffer appears again in the input data, the buffer indicates the start position and length of the character string. A set of indices is a codeword for the character string. The buffer is updated so that the latest data that has been compressed is always stored. When the block at the head of the cluster is compressed, an initial value of the buffer is used. When the state is set to a predetermined setting, and when the block other than the head of the cluster is compressed, the initial state of the buffer is the buffer at the end of the compression of the block at the head of the cluster to which the block belongs. 2. The data compression control method according to claim 1, wherein: 5. A dictionary for associating an index with a character string in input data which has already been compressed as said model, and when a character string in said dictionary appears again, an index of said character string in said dictionary Is a code word for the character string, the dictionary is updated by assigning and registering an index to a character string newly appearing in the input data, and when the block at the head of the cluster is compressed, The initial state of the dictionary is set to a predetermined setting. When the blocks other than the head of the cluster are compressed, the initial state of the dictionary is set when the compression of the head of the cluster to which the block belongs is completed. 2. The data compression control method according to claim 1, wherein the state of the dictionary is set as follows. 6. A structure in which a set of contexts, which are character strings appearing in input data, and a frequency of appearance of each character in input data under said context are used as said model, Encoding each character in the input data based on the appearance frequency, updating the set of contexts by adding a character string newly appearing in the input data to the set of contexts,
Updating the frequency of appearance each time a character appears under the context in the input data; and compressing the block at the beginning of the cluster, the set of contexts and the appearance of each character under the context. The initial state of the frequency is set to a predetermined setting. When the block other than the head of the cluster is compressed, the context obtained when the compression of the block at the head of the cluster to which the block belongs is completed. 2. The data compression control method according to claim 1, wherein an initial state of the appearance frequency of each character under a set and a context is set.