JP2002232293A - Device and method for compressing image data, program for computer to perform image data compressing method and computer readable recording medium recording program for computer to perform image data compressing method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an image data compressor or the like capable of improving compression of a DC value and improving entire compression performance by encoding the DC value so that the code amount of the DC value can be reduced. SOLUTION: In an image data compressor 1, the image data of an original image are inputted by an image data input part 3, a DC image is prepared by a DC image preparing part 5, and the vector quantization of mean value separate type block encoding concerning the DC value is performed by a DC vector quantizing part 6. The encoded DC vector is decoded by a DC vector decoding part 13, and the vector quantization of mean value separate type block encoding concerning an AC value is performed by an AC vector quantizing part 16.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to an image data compression apparatus, an image data compression method, a program for causing a computer to execute the image data compression method, and a program for causing a computer to execute the image data compression method. With respect to the computer-readable recording medium, especially, the image data is divided into a plurality of pixel blocks,
An image data compression device that encodes a DC value representing an average of pixel values for each block, an image data compression method, a program for causing a computer to execute the image data compression method, and an image data compression method executed by a computer The present invention relates to a computer-readable recording medium that stores a program for causing a computer to execute the program.

[0002]

2. Description of the Related Art A vector quantization method is one of the elemental techniques in a data compression technique for reducing the amount of image data, which is original data usable by a computer, by encoding.

In the vector quantization method, an appropriate number of data samples from the original data are collectively treated as a vector.
For the encoding target vector, an optimal vector whose error from the encoding target vector is within an allowable value is retrieved from a set of vectors called a dictionary prepared in advance, and the encoding target vector is represented by a number in the dictionary. This is a method for reducing the amount of information.

In order to improve the approximation accuracy and reduce the amount of data in the dictionary in order to reduce the allowable value of the error, the encoding target vector is represented by additional information called a multiplier coefficient in addition to the number in the dictionary. But international publication number W
No. 00/02393.

FIG. 7 is a diagram showing the internal configuration of a conventional image data compression device, and FIG. 8 is a flowchart for explaining the operation of the image data compression device of FIG.

The image data compression device 51 includes an image data input unit 53, a DC image creation unit 55, and a DPCM processing unit 5.
7, a run length encoding unit 59, and a Huffman encoding unit 61
, Run-length decoding unit 63 and Huffman decoding unit 65
, A DC image decoding unit 67 and an AC vector creation unit 69
, AC dictionary holding unit 71, AC vector encoding unit 73
And

[0007] The image data input unit 53 inputs image data as original data to the DC image generation unit 55 and the AC vector generation unit 69. The DC image creation unit 55 creates a DC image and inputs the DC image data to the DPCM processing unit 57.

[0008] The DPCM processing unit 57 performs a DPCM encoding process described below on the DC image, and inputs the processing result to the run-length encoding unit 59 and the Huffman encoding unit 61.
The run-length encoding unit 59 performs a later-described run-length encoding process, and the Huffman encoding unit 61 performs a later-described Huffman encoding process, and outputs a DC value code.

The output DC value code is input to a run-length decoding unit 63 and a Huffman decoding unit 65 for generating an AC value code. Run-length decoding unit 6
3 performs a run-length decoding process on the input DC value code, and the Huffman decoding unit 65 outputs
Huffman decoding is performed on the value code.

[0010] The processing result of the run-length decoding unit 63 and the processing result of the Huffman decoding unit 65 are input to a DC image decoding unit 67. The DC image decoding unit 67 decodes the DC image to generate and decode the DC image. The data is input to the AC vector creation unit 69.

The AC vector creation section 69 creates an AC vector from the input original data and DC image data, and inputs the created AC vector to the AC vector encoding section 73. The AC vector encoding unit 73 searches the set of vectors, which are the AC dictionaries held in the AC dictionary holding unit 71 required for the vector quantization, for an optimum vector whose error from the AC vector is within an allowable value. , A expressed by a number and a multiplier coefficient
The C value code is obtained and functions as an output unit, and outputs a DC value, a vector number, and a multiplier coefficient as a code of an AC value.

The operation will be described with reference to FIG.
In step ST1, image data as original data is read. An RGB image is converted to a YUV image and read. Y is luminance data,
U and V correspond to color difference data, and U and V are down-sampled using the average luminance of two horizontal pixels.

In step ST2, DC values are separated for each image block, and in step ST3,
A DC image is created. This will be further described below.

First, 4 ×
Block data T _{J, I of} four pixels is extracted. A DC value M, which is an average of 16 pixel data included in the block data T _{J, I} is obtained. The DC value M is stored in the storage position DC _{J, I} of the DC image.

In step ST4, DPC is added to the DC value.
M processing is performed, run length and Huffman coding are performed, and in step ST5, a code of a DC value is output.

The DPCM processing will be described. When J rows of the DC image, the DC value of the column I D _J, and _I, D _J, the prediction value D _'J, the _I D' of _{I J, I = (D J} , I-1 + D J-1 _{, I} ) / 2. Let the prediction error ΔD _{J, I} be ΔD _{J, I} = D _{J, I} −D ′ _{J, I}
Ask by

Here, since the DC value is an average value of 16 pixels included in the image block, the change in the DC value is very small as compared with the change in the pixel value, so that the prediction error ΔD
_{J and I} are expected to be small values. If the value is small, the number of bits allocated to the code can be reduced. Therefore, in order to further reduce the code amount, the prediction error ΔD _{J, I} is integer-divided by the quantization coefficient Q according to the target image quality. Is the quantization prediction error Δd _{J, I} , and Δd _{J, I} = (in
t) (ΔD _{J, I} / Q). Here, (i
nt) (a) indicates a value obtained by rounding the real number a. The quantization coefficient Q is a coefficient having a small value when high image quality is required.

In many cases _{, the} quantization prediction error Δd _{J, I} has a value of zero. In step ST4, run-length encoding is performed on zero and Huffman encoding is performed on non-zero. , To reduce the code amount.

[0019] Run-length coding is a simple technique used to compress a sequence of the same symbols in a sequence of data, where the compression is average when a sequence occurs.

FIG. 9 is a diagram showing an example of a Huffman tree for explaining Huffman coding. A Huffman tree is created by the following five procedures.

The first is a procedure for extracting two nodes having the smallest weight from the free node list.

The second is a procedure in which a parent node is created for the two extracted nodes and the weight sum of the two child nodes is set as the weight for the parent node.

The third is a procedure for adding the parent node of the free node list and deleting two child nodes from the list.

Fourth, bit 0 is assigned to one of the child nodes.
, And bit 1 to the other, to create a subtree with the parent node.

Fifth, the first to fourth procedures are repeated until the number of nodes remaining in the free node list becomes one, and the last remaining node is assigned to the root of the tree.

The Huffman tree shown in FIG. 9 is obtained by the above first to fifth procedures. When viewed from the root along each symbol, the sign of the symbol A is “0” and the symbol B is , The symbol C is “101”, the symbol D is “110”, and the symbol E is “111”.

In step ST6 of FIG. 8, the DC value code is decoded, and the DC image is decoded. This decryption
Basically, the processing is performed by the reverse of the processing from step ST3 to step ST5.

In step ST7, a calculation process for obtaining an AC vector is performed for each image block, and in step ST8, all the AC vectors are obtained. The AC vector is a 16-dimensional residual vector <d> corresponding to a 4 × 4 image block, and is represented by <d> = T _{J, I} −D _{J, I.} Here, <> represents a vector. In addition, 8 bits are allocated to each element of the AC vector <d>.

Then, in step ST9, the vector quantization of the AC vector is performed, and the encoding of the AC value is performed. The encoding performed by vector quantization using a prepared dictionary performed here will be described in detail below.

First, the Gram-Schmidt orthogonalization will be described. The orthogonalization of the Gram-Schmidt configuration, one of the underlying n-dimensional inner-product space _{V {v 1, ···, v} n} from the V orthonormal basis _{{v '1, ···, v} ' n} the How to

The first basis vector <v ₁ > (where <u
₁ >), the first normalized base vector <v ′ ₁ > is a unit vector, and is expressed as in the following equation (1). <V ′ ₁ > = <v ₁ > / ‖ <v ₁ > ‖ = α ₁₁ <u ₁ > (1)

Next, assuming that the second base vector <u ₂ > is extracted, the first normalized base vector <v ′ ₁ >
A second orthogonal basis vector <v ₂ > that is orthogonal to
Equation (2) can be used. <V ₂ > = <u ₂ > + k <v ′ ₁ > (2)

Let (<a>, <b>) be the inner product of the vector
Notation, (<v_Two>, <V ' ₁>) = 0,
Equation (3) is obtained. (<V_Two>, <V '₁>) = (<U_Two> + K <v '₁>, <V '₁>) = (<U_Two>, <V '₁>) + K (<v ′₁>, <V '₁>) = (<U_Two>, <V '₁>) + K = 0 (3)

The scalar coefficient k at this time is k = − (<
u ₂ >, <v ′ ₁ >). By substituting k into Equation (2), Equation (4) is obtained. <V ₂ > = <u ₂ > − (<u ₂ >, <v ′ ₁ >) <v ′ ₁ > (4)

Since the second normalized basis vector <v ′ ₂ > is also a unit vector, it is expressed by the following equation (5). <V ′ ₂ > = <v ₂ > / ‖ <v ₂ > ‖ = α ₂₁ <u ₁ > + α ₂₂ <u ₂ > (5)

Similarly, in general, the n-th normalized basis vector <v ′ _n > is given by the following equation (6). However, the (6)
Α _ni in the equation is an expansion coefficient.

[0037]

(Equation 1)

Based on the above idea, the encoding target AC
Consider the vector <AC>. AC vector to be encoded <
AC> is a 16-dimensional vector, and has 16 AC values ordered in correspondence with each pixel as elements.

First, <AC> is converted to a dictionary vector <ACV ₁
Try to approximate using>. <AC> and dictionary vector <A
The magnitude of the difference vector between CV _1> alpha ₁ multiplied by a multiplier factor alpha ₁ to _{<ACV 1>{<AC>} -α 1 <ACV 1>} is minimized, alpha ₁ and <ACV _1> Difference vector {<AC> −α ₁ <
ACV ₁ >} is orthogonal. Accordingly, Expression (7) representing that the inner product is 0 is obtained. ({<AC> −α ₁ <ACV ₁ >}, α ₁ <ACV ₁ >) = 0 (7)

Α₁Solving for α₁= (<AC>, <
ACV₁>) / (<ACV₁, ACV ₁> Is obtained. So
Then, the encoding target vector <AC> is expressed by the following equation (8).
Approximation. <AC '> = α₁<ACV₁> ... (8)

Next, an approximation error vector <AC ₁ > based on the dictionary vector <ACV ₁ > is obtained by using the equation (8).
It is expressed as in equation (9). <AC ₁ > = <AC> − <AC ′> = <AC> −α ₁ <ACV ₁ > (9)

Then, in step ST10 in FIG. 8, the vector number and the multiplier coefficient in the dictionary in which the approximate error vector is the smallest are output as the sign of the AC value.

[0043]

Since the image data Y, U, and V are each represented by 8 bits, the code amount of each DC value is, as described above, 1 byte (16 bytes) for 16 pixels. 8 bits). Here, assuming that Y uses data of all pixels and U and V use only one data for every four pixels, that is, Y: U: V =
Assuming that the ratio is 4: 1: 1, the data of one pixel includes Y,
Since one piece of data is required for each of U and V, the sum of the data amount of one pixel is U and V, which are １ ／, where Y is 1, and 1 + 1/4 + ／ =
The data amount is 1.5.

Therefore, in addition to the DC value for Y of 16 pixels being 8 bits, the total data amount considering U and V is 12 bits. If this is converted into the data amount per pixel, it becomes 12/16 = 0.75 bits.

Further, a DPCM for encoding a DC value
According to the data compression by Huffman coding, although there is a variation depending on the image when the data amount is 1/4 and the data amount is 0.8 times as large, according to the experimental results for a large number of image data, It has been found that the data volume can be reduced to about half if it were done.

Therefore, when the data amount of the converted DC value per pixel described above is compressed by DPCM and Huffman coding, the average is 0.75 × 0.5 =
The code amount of 0.375 is reduced to about 0.4 bits.

On the other hand, International Publication Number WO00 / 02393
When examining the code amount per pixel,
When the code amount obtained by encoding the AC value is 0.4 bits, D
It has been found that the code amount obtained by coding the C value is also about 0.4 bits.

However, even if the code amount of the AC value is reduced to 0.2 bits by lowering the target image quality in order to improve the compression performance, the code amount of the DC value remains at 0.4 bits, so that the overall Compression performance does not improve up to twice,
It stays about 1.3 times. If the code amount of the AC value becomes 0 bits, the entire compression performance doubles for the first time.

That is, in the prior art, even if the image quality is reduced and the compression performance is improved, only the code amount of the AC value is reduced, but the code amount of the DC value is not reduced. The improvement in compression performance as a whole could not be obtained as expected.

Therefore, an object of the present invention is to improve the compression performance of a DC value by encoding a DC value capable of reducing the code amount of the DC value, and thereby to improve the overall compression performance. Provided are an image data compression apparatus, an image data compression method, a program for causing a computer to execute the image data compression method, and a computer-readable recording medium that stores a program for causing a computer to execute the image data compression method. It is in.

[0051]

According to a first aspect of the present invention, there is provided an image data compression apparatus for encoding and compressing image data, wherein the image data is divided into a plurality of pixel blocks, and an average of pixel values of each block is obtained. DC image creating means for creating a DC image composed of DC values representing the DC value, and an average value of a plurality of DC values of the DC image created by the DC image creating means is determined. From the DC values for which the average value is determined, DC vector quantization means for subtracting an average value, creating a DC vector having a plurality of values obtained by the subtraction as elements, and vector-quantizing and encoding the created DC vector. is there.

Therefore, in order to encode a DC value representing the average of the pixel values for each block when the image data is divided into blocks of a plurality of pixels, a plurality of blocks are used instead of the conventional Huffman coding or the like. The average value of the DC values is obtained, the average value is subtracted from each of the DC values for which the average value is obtained, a DC vector having a plurality of values obtained by the subtraction as elements is created, and the created DC vector is vector-quantized. And encode it. Huffman coding is a bit-by-bit process, and although it is a process with a heavy computational load in a general CPU, it can reduce the code amount by about half, so it is used in coding of DC values of image data such as JPEG. However, according to the present invention, it is not necessary to use the Huffman coding process that requires a long processing time.

According to a second aspect of the present invention, in the first aspect, the number of DC values required for obtaining the average value is an integer multiple of twice or more the number of elements of the DC vector. It is assumed that.

Therefore, in order to generate a DC vector, the number of DC values that is an integer multiple of twice or more the number of elements of the DC vector is required as an average value in the mean value separation type block coding method. And such an average is D
The variation is smaller than the variation of the C value. In particular, the larger the value of the integral multiple is, the more the average value used for encoding the DC vector is commonly used for many DC vectors (DC vector elements).

The invention according to claim 3 is the invention according to claim 1 or 2
Wherein the DC vector quantization means includes a DC vector creating unit for creating the DC vector, and a DC dictionary holding a DC dictionary which is a set of vectors for encoding the DC vector created by the DC vector creating unit. The vector between which the error between the dictionary holding unit for use and the DC vector created by the DC vector creation unit is within an allowable value
And a DC vector encoding means for retrieving from a vector in the DC dictionary held by the dictionary holding unit and encoding the DC vector.

Therefore, a vector whose error from the DC vector is within an allowable value is searched from the vector in the DC dictionary, and the DC value is encoded by at least the DC vector number specifying the searched vector.

According to a fourth aspect of the present invention, the DC vector encoding means according to the third aspect, wherein the DC vector encoding means converts the vector having an error within a tolerance from the DC vector created by the DC vector creating section into the DC dictionary. The vector in the DC dictionary held by the holding unit is searched by multiplying the vector by a multiplier coefficient, and the vector corresponding to the DC vector number and the DC vector number for specifying each vector of the vector set as the DC dictionary. The DC vector is encoded by a multiplier coefficient.

Therefore, the DC value is encoded by the corresponding multiplier coefficient which can improve the approximation accuracy and reduce the number of dictionary vectors, in addition to the DC vector number specifying the searched vector.

According to a fifth aspect of the present invention, in the fourth aspect, the output means for outputting the DC vector number, a multiplier coefficient corresponding to the DC vector number, and the average value as a sign of a DC value of the image data. It is provided with.

Therefore, as a code of a DC value for compressing image data, a DC vector number, a multiplier coefficient corresponding to the DC vector number, and the average value are output.
The output code can be decoded.

The invention according to claim 6 is the invention according to claims 1 to 5
In any one of the above, the DC value is subtracted from each pixel value of the image data, and an AC vector having a plurality of values obtained by the subtraction as elements is created.
It is provided with an AC vector quantization means for vector-quantizing and encoding a vector.

Therefore, not only the DC value but also the AC value is encoded by vector quantization, and the unified encoding of the DC value and the AC value is performed on the image data.

According to a seventh aspect of the present invention, in the sixth aspect, a DC vector decoding means for decoding the DC vector encoded by the DC vector quantization means, and a DC image from the DC vector decoded by the DC vector decoding means. DC image decoding means for decoding the DC image, wherein the AC vector quantization means uses the DC value of the DC image decoded by the DC image decoding means as the DC value.

Therefore, when the AC vector is created, the AC vector is created using the DC value of the DC image decoded from the DC vector instead of the DC image created from the image data of the original image. AC that takes into account decoding the original image as accurately as possible
Vector encoding can be performed.

The invention according to claim 8 is the invention according to claim 6 or 7
In the above, the AC vector quantizing means includes an AC vector creating means for creating the AC vector and an AC dictionary holding an AC dictionary which is a set of vectors for encoding the AC vector created by the AC vector creating means. Dictionary holding unit, and the AC created by the AC vector creating unit.
AC vector coding means for searching a vector in the AC dictionary held by the AC dictionary holding unit for a vector whose error between the vector and the vector is within an allowable value, and coding the AC vector. It is.

Therefore, a vector whose error from the AC vector is within an allowable value is searched from the vector in the AC dictionary, and the AC value (AC vector) is at least identified by the AC vector number specifying the searched vector. Be transformed into

According to a ninth aspect of the present invention, in the ninth aspect of the present invention, the AC vector encoding means according to claim 8, wherein the vector having an error between the AC vector generated by the AC vector generating unit and the AC vector is within an allowable value. The vector in the AC dictionary held by the holding unit is searched by multiplying the vector by a multiplier coefficient, and the vector corresponding to the AC vector number and the AC vector number for specifying each vector of the set of vectors that are the AC dictionary. The AC vector is encoded by a multiplier coefficient.

Therefore, the AC value is encoded by the corresponding multiplier coefficient that can improve the approximation accuracy and reduce the number of dictionary vectors, in addition to the AC vector number that specifies the searched vector.

According to a tenth aspect of the present invention, in the ninth aspect, output means for outputting the AC vector number, a multiplier coefficient corresponding to the AC vector number, and the DC value as a code of the image data. Things.

Therefore, as the AC value code for compressing the image data, the AC vector number, the multiplier coefficient corresponding to the AC vector number, and the DC value are output, and the output code can be decoded.

According to an eleventh aspect of the present invention, in the image data compression method for encoding and compressing image data, the image data is divided into blocks of a plurality of pixels, and comprises a DC value representing an average of pixel values for each block. A first step of creating a DC image, obtaining an average value of a plurality of DC values of the DC image created in the first step, and subtracting the average value from each DC value for which the average value has been obtained. ,
A second step of creating a DC vector having a plurality of values obtained by the subtraction as elements, vector-quantizing and encoding the created DC vector, and the DC step from each pixel value of the image data. Subtracting the values, creating an AC vector having a plurality of values obtained by the subtraction as elements,
A third step of vector-quantizing and coding the generated AC vector.

Therefore, in order to encode a DC value representing the average of the pixel values for each block when the image data is divided into blocks of a plurality of pixels, a plurality of blocks are used instead of the conventional Huffman coding or the like. The average value of the DC values is obtained, the average value is subtracted from each DC value for which the average value is obtained, a DC vector having a plurality of values obtained by the subtraction as elements is created, and the created DC vector is vector Quantize and encode. Huffman coding is a bit-by-bit process, and although it is a process with a heavy computational load in a general CPU, it can reduce the code amount by about half, so it is used in coding of DC values of image data such as JPEG. However, according to the present invention, it is not necessary to use the Huffman coding process that requires a long processing time.

Further, a DC value is subtracted from each pixel value of the image data, and an AC having a plurality of values obtained by the subtraction as elements is obtained.
A vector is created, and the created AC vector is also vector-quantized and encoded. Therefore, not only the encoding of the DC value but also the encoding of the AC value is vector-quantized, so that the original image data can be encoded by the unified encoding method.

According to a twelfth aspect of the present invention, in the eleventh aspect, the number of DC values required for obtaining the average value is an integer multiple of twice or more the number of elements of the DC vector. It is assumed that.

According to a thirteenth aspect of the present invention, in the eleventh or twelfth aspect, a DC dictionary as a set of vectors for encoding the DC vector created in the first step is prepared in advance. In the second step, a vector whose error between the DC vector created in the first step and the DC vector is within an allowable value is converted to the DC vector.
And searching the vector in the dictionary for encoding and encoding the DC vector.

According to a fourteenth aspect of the present invention, in the thirteenth aspect, the second step uses the vector for the DC whose error between the DC vector and the DC vector created in the first step is within an allowable value. The vector in the dictionary is searched by multiplying it by a multiplier coefficient, and the DC vector is identified by a DC vector number specifying each vector of the set of vectors as the DC dictionary and a multiplier coefficient corresponding to the DC vector number. Is encoded.

According to a fifteenth aspect, in the fourteenth aspect, a fourth step of outputting the DC vector number, a multiplier coefficient corresponding to the DC vector number, and the average value as a code of the image data. Including.

According to a sixteenth aspect of the present invention, in any one of the eleventh to fifteenth aspects, the fifth step of decoding the DC vector encoded in the second step and the step of decoding the DC vector encoded in the fifth step are performed. And a sixth step of decoding a DC image from the obtained DC vector. The third step includes a step of using a DC value of the DC image decoded in the sixth step as the DC value.

Therefore, the invention of claim 12 is based on claim 2
The same operation as the operation of the invention of the third aspect is obtained. The invention of the thirteenth aspect obtains the same operation as the operation of the third aspect of the invention, and the invention of the fourteenth aspect has the same operation as the operation of the fourth aspect of the invention. The invention of claim 15 has the same function as the function of the invention of claim 5, and the invention of claim 16 has the function of the invention of claim 7.

The invention according to claim 17 is a program for causing a computer to execute the image data compression method according to any one of claims 11 to 16.

The invention according to claim 18 is a computer-readable recording medium on which the program according to claim 17 is recorded.

[0082]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Preferred embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a schematic block diagram of an image data compression device according to an embodiment of the present invention.

The image data compression device 1 includes an image data input unit 3, a DC image creation unit 5, and a DC vector creation unit 7.
, DC vector encoding unit 9, DC dictionary holding unit 11
, DC vector decoding unit 13 and DC image decoding unit 15
, AC vector creation unit 17, AC dictionary holding unit 19
And an AC vector encoding unit 21. A DC vector quantization unit 6 is constituted by the DC vector creation unit 7, the DC vector encoding unit 9, and the DC dictionary holding unit 11, and AC
The vector creation unit 17, the AC dictionary holding unit 19, and the AC vector encoding unit 21 constitute an AC vector quantization unit 16.

The image data compression apparatus 1 is different from the conventional image data compression apparatus 51 shown in FIG. 7 in that it includes a DC vector creation section 7, a DC vector encoding section 9, and a DC dictionary holding section 11. To vector quantize the DC value,
The code amount of the DC value is reduced.

Image data input unit 3 and DC image creation unit 5
Are the image data input unit 53 and the DC image creation unit 5 of FIG.
5, the DC image data is input from the DC image generation unit 5 to the DC vector generation unit 7. The DC vector creation unit 7 creates a DC vector and inputs it to the DC vector encoding unit 9.

The DC dictionary holding unit 11 holds a dictionary which is a set of vectors for encoding DC vectors. The DC vector encoding unit 9 includes a DC dictionary holding unit 11
From the set of vectors, a search is made for an optimal vector whose error from the DC vector is within an allowable value, a DC value code represented by a number and a multiplier coefficient is obtained, and the output unit functions as an output unit. Vector number and multiplier coefficient are D
Output as the sign of the C value. The output DC value code is
It is input to the DC vector decoding unit 13. The DC vector decoding unit 13 decodes the DC vector from the DC value code and
It is input to the C image decoding unit 15.

The DC image decoding unit 15 decodes a DC image from a DC vector and inputs DC image data to an AC vector creation unit 17. The AC vector creation unit 17, the AC dictionary holding unit 19, and the AC vector encoding unit 21 correspond to the AC vector creation unit 69, the AC dictionary holding unit 71, and the A
The operation is similar to that of the C vector encoding unit 73, and the description is omitted.

FIG. 2 is a flowchart for explaining the outline of the process of creating the DC dictionary held by the DC dictionary holding unit of FIG. 1, and FIG. 3 further shows the process of the flowchart of FIG. FIG. 4 is a flowchart showing details, and FIG.
It is the flowchart which showed the process of S13, SS14, and SS15 in more detail.

In step S1 of FIG. 2, a DC vector is read. Specifically, as shown in step SS1 of FIG. 3, the DC vector separated by the average value is read into the array Vin [num]. At this time, the norm of each vector is normalized to 1, and when the first element of the vector is a negative number, the signs of all elements are changed to make the first element a positive number. Here, the number of source vectors, which is a set of DC vectors collected from a plurality of image data, is
It is represented by num.

FIG. 5 shows the pixels, DC values and D of the image data.
FIG. 5 is a diagram for explaining a relationship between C vectors.
FIG. 5A is a diagram showing image data, and FIG.
It is the figure which showed the C image.

The DC value is created by an average value of 16 pixel values, and the DC vector is created by, for example, 16 DC values in the range 31. The DC vector is obtained by separating the average values, and is obtained by subtracting the average value of the 64 DC values from each of the 16 DC values. Therefore, the number of DC vectors collected from one image data is 256 pixels.
Divided by.

However, although the image data is recorded for each of the R, G, and B color components, it is converted into Y, U, and V components and used in encoding the image data. For U and V, since only one data is used for four data, the total number of vectors is 2 pixels of the image.
A value obtained by multiplying the value obtained by dividing by 56 by 1.5 is obtained.

Here, the image data used when creating a dictionary is not always the same size. For example, if the image data is 640 in length and 480 in width,
× 480/256 × 1.5 = 1800 vectors are collected. However, in the actual processing of the present invention, vectors having a norm of less than 16 are not collected, so that the number is less than 1800. Further, for example, in the case of image data having a height of 1600 and a width of 1200, the maximum is 1600 × 1200/256 × 1.5 = 111250.
Vectors are collected. Therefore, if source images are collected using 50 images of 640 height and 480 width and 50 images of 1600 height and 1200 width, a maximum of 1800 × 50 + 1250 × 50 = 65250
Zero vectors will be collected.

Step S2 in FIG. 2 and step S in FIG.
In S2, errst representing the accumulated error with respect to the current dictionary vector is initialized to errst = 0, and the number of dictionary vectors is set to n.

In vector quantization, a vector included in the created dictionary is used to approximate a vector of arbitrary image data. Therefore, when the number of vectors included in the dictionary is n, n vectors are included. It is necessary to make vector approximation performance for various images as high as possible. Therefore, after collecting various vectors with the number of source vectors num being about 100 times the value of n, the n vectors having the highest approximation performance to the entire collected vectors are selected.

Step S3 in FIG. 2 and step S in FIG.
In S3, the dictionary is initialized. That is, different vectors are copied from the dictionary array SGorg [0] to SGorg [n-1], and n vectors are set.

In step S4 of FIG. 2, when each DC vector is approximated by a dictionary vector, the number of the dictionary vector that minimizes the error is determined, and the errors are accumulated. This process is a process from step SS4 to step SS11 in FIG.

In step SS4, errnew, which is the total error when the current dictionary is improved, is set to errn.
ew = 0, and the index I of the array Vin is initialized to I = 0.

In step SS5, the array SGorg
Is initialized to K = 0, the variable er representing the minimum error is initialized to a very large value of er = 1e30, and the array Vin [I] is copied to the vector D. Here, er is the smallest error value recorded up to the present time by the processing of step SS8 described later, and 1e3
0 represents 1 × 10 ＾ 30 (1 × 10 to the 30th power).

In step SS6, the array SGorg [K] is copied to the vector V, an approximate error vector E obtained when the vector D is approximated by the vector V is obtained, and the square norm of the vector E is substituted for the minimum error errnow. Here, alp is alp = (<D>, <V>) /
(<V>, <V>), the expression is transformed into alp = (<D>,
<V>). This is because the norm of each vector is normalized to 1 at the stage of reading the source vector,
This is because (<V>, <V>) = 1. And K
= K + 1 and proceed to the next dictionary number.

In step SS7, errnow and er are compared. If errnow <er, the process proceeds to step SS8, and if ernow = er, the process proceeds to step SS9.

In step SS8, er = errn
The minimum error is updated as ow, and grp [I] = K−1 to indicate that the vector Vin [I] has the minimum error by SGorg [K−1].

In step SS9, it is determined whether K is smaller than n. If K <n, a loop process is performed to return to step SS6 to check the next dictionary array vector, and if K = n, Proceed to step SS10.

Next, steps SS7, 8, and 9 will be described. grp [I] is the source vector Vin [I]
Indicates that the error is minimized by the grp [I] -th vector of the dictionary vector. e
If rrnow is smaller than er, the current dictionary vector has the smallest error so far. However, since K is advanced by one in step SS6, grp [I] = K− Let it be 1. While K <n, the dictionary vector that minimizes the error is continuously searched.

At step SS10, the array SGo
When the survey is completed up to the end of g, er contains the array S
Since the value in which the error is minimized by the approximation in Gorg is stored, er is added to a variable errnew representing the total error in the entire source vector, and updating is performed. Then, the variable I of the array Vin is I = I
It is updated to +1.

In step SS11, it is determined whether I is smaller than num. If I <num, the process returns to step SS5 to check the next source vector.
When I reaches num, the process proceeds to step SS12.

Step S5 in FIG. 2 and step S5 in FIG.
In S12, since the investigation for all the source vectors has been completed, eps = | errst-er
rnew | / errnew is determined. Then, in preparation for the next improvement process, the current error is updated with errst = errnew.

Step S6 in FIG. 2 and step S in FIG.
In S13, it is determined whether esp is a value smaller than 0.001. If esp <0.001, the process proceeds to step SS15, and if esp ≧ 0.001, the process proceeds to step SS14.

First, when errst = errornew, it is indicated that no further improvement is possible, so it is time to terminate the dictionary improvement processing. However, continuing the dictionary improvement processing to this stage takes a very long time, and thus, in order to speed up the dictionary improvement processing, eps = | errst-err
The change rate of the error is calculated as nnew | / errnew,
At the stage where esp <0.001 is satisfied, the dictionary is discontinued.

Note that the value of 0.001 indicates that the rate of change of the error is extremely small, but is not limited to this, and other values such as 0.01 and 0.0001 can be used. The dictionary improvement processing may be completed in a shorter time as the numerical value increases.

Step S7 in FIG. 2 and step S in FIG.
In S14, a vector in which the number of the dictionary vector with the minimum error in each DC vector is M is added to M representing the vector number of the dictionary from 0 to n−1, that is, grp [L] = M A certain vector Vin [L] is accumulated, and the norm of the vector is normalized to 1. And
The dictionary vector is updated as a new SGorg [M].

Step S8 in FIG. 2 and step S in FIG.
In S15, the norm of the dictionary vector SGorg [M] is normalized to 100 for M representing the dictionary vector number from 0 to n−1, and is converted to an integer.

The processing after step S6 in FIG. 2 and step SS13 in FIG. 3 will be further described with reference to FIG.

In step SS141, dictionary number M
Are initialized to M = 0. In step SS142, the vector v is initialized to v = 0. Step SS
At 143, the source vector number L is initialized to L = 0.

In step SS144, grp
It is determined whether [L] = M. If grp [L] = M, the process proceeds to step SS145, and the process proceeds to step SS14.
In order to calculate the vector sum at step 5, v is Vin
After adding [L] to make v = v + Vin [L], the process proceeds to step SS146. If grp [L] ≠ M, the processing for vector v is not performed and step SS146 is not performed.
Proceed to 146. In step SS146, the source vector number L is updated to L = L + 1.

In step SS147, L is num
It is determined whether it is smaller than. If L <num, the process returns to step SS144. If L = num, the process proceeds to step SS148.

In step SS148, v = v / |
v | to normalize the norm of the vector v to 1, SGo
The dictionary is updated with the vector normalized by rg [M] = v, and the dictionary number is advanced to the next by M = M + 1.

In step SS149, it is determined whether or not M <n. If M <n, the process returns to step SS142. If M = n, the dictionary update processing is terminated.
The process returns to step S4 of FIG. Thereby, step S7 of FIG. 2 and step SS1 of FIG.
4 is completed.

In step SS151 in FIG. 4, the dictionary number M is initialized to M = 0. In step SS152, SGorg [M] is copied to the vector v,
The norm is normalized to 100 by v = 100 × v / | v |, the vector element is converted to an integer by v = (int) v, and the dictionary vector SGorg [M] = v is updated.

In step SS153, M is M = M +
1 and it is determined in step SS154 whether M is smaller than n. If M <n, the process returns to step SS152, and if M = n, the process ends. Thereby, the processing of step S8 of FIG. 2 and step SS15 of FIG. 3 ends.

By the way, when the vector quantization of the DC vector described later is performed, the vector number and the multiplier coefficient are obtained. In the experimental result when the norm of the dictionary vector is normalized to 100, the multiplier coefficient is 0.01 to 1
It was found to be in the range of 0 values. In decibel values,
The range is from −40 to 20 dB, and the dynamic range of the decibel value is 20 − (− 40) = 60 dB.
Furthermore, since the multiplier coefficient may be positive or negative, it has been found that one bit is required for the sign to indicate positive or negative. Eventually, it was found that the DC vector is represented by the vector number, the sign of the multiplier coefficient, and the number of the multiplier coefficient.

When the vector number, the sign of the multiplier coefficient and the multiplier coefficient number are 16 bits, if the number of vectors in the dictionary is 1024, 2 to the 10th power is 1024, so the vector number is 10 bits. Is required.
Therefore, the code amount that can be used for the multiplier coefficient number is 32 types of 5 bits by subtracting 10 bits from 16 bits and further subtracting 1 bit of the sign representing the sign. 60d
When the dynamic range of B is divided into 32 equal parts, the interval of the multiplier coefficient is 60/31 = 1.935 dB.

Similarly, the number of dictionary vectors is set to 512 (2
, The multiplier coefficient number is 16-9.
64 types of -1 = 6 bits can be used. 60d
When the dynamic range of B is divided into 64 equal parts, the interval of the multiplier coefficient is 60/63 = 0.954 dB.

In the case of using a general multiplier coefficient, when n bits are used for the multiplier coefficient number, 2 n multiplier coefficients can be used, so that the multiplier coefficient is expressed by the following equation (10). You.

(Equation 2)

FIG. 6 is a flowchart for explaining the operation of the image data compression apparatus of FIG.

At step TS1, image data as an original image is read. At step TS2, the image data is divided into square areas (blocks) to separate DC values for each image block. The DC value is an average value for each image block, and in step TS3,
A DC image is created.

In step TS4, the average value m of the 64 DC values processed for each of the 64 DC values is calculated.
A value obtained by subtracting m from the four DC values is calculated, four DC vectors each having 16 values are created, and each of the DC vectors is subjected to vector quantization.

The creation of four DC vectors will be briefly described with reference to FIG. Vector elements are obtained from the DC image 25 of FIG. 5B by a total of 16 number-ordered DC values of 4 pixels in the horizontal direction and the vertical direction of the range 31. Take four vectors in the horizontal direction,
Calculate the average value m for the 64 DC values. M is subtracted from each of the elements of the four vectors, and four vectors having the average value separated are created. Note that the average value is represented by 8 bits.

Next, vector quantization is performed on a vector (encoding target vector) separated by the average value using a dictionary prepared in advance. By reducing the change in the DC value, the average value of the four vectors is subtracted from each vector, and vector quantization is performed on the vector obtained by separating the average value, thereby reducing the code amount of the DC value. Is being planned.

If the encoding target vector is represented by <DC>,
Using the dictionary vector <DCV1>, <DC> becomes <D
C '> = α1 <DCV1>.
Here, α1 is a multiplier coefficient for the dictionary vector <DCV1>, and when the inner product of the vector <a> and the vector <b> is represented by (<a>, <b>), α1 = (<DC>, <D
CV1>) / (<DCV1>, <DCV1>).

The approximate error vector <DC1> based on the dictionary vector <DCV1> is expressed as <DC1> = <DC> − <DC ′
> = <DC> −α1 <DCV1>. The approximation error err is err = (<DC1>, <DC1
>).

By examining all dictionary vectors, an approximation error e
The number of the dictionary vector with the smallest rr is No, and if the sign of the multiplier coefficient is positive, sign = 0, and if negative, sign
= 1. The conversion of the multiplier coefficient into the number B in Table G [i] is performed according to Expression (11).

[0133]

(Equation 3)

In the proviso of the equation (11), A = 20 l
og10 (A) +40 is an expression meaning that the value of A is updated using the current variable A (the absolute value of the multiplier coefficient).

In step TS5, a DC value code is output. The code for the DC value is output in the form of one average value, four vector numbers, and four multiplier coefficients for 64 DC values.

Thereafter, in step TS6, the DC value code is decoded to decode the DC image. In step TS7, an AC vector is calculated for each image block. In step TS8, an AC vector is created. Quantization is performed, and an AC value code is output in step TS10. The creation of the AC vector is performed by subtracting the DC value from each pixel value of the original image block, and the quantization of the AC vector is performed using an AC dictionary.

A description will now be given of the fact that the code amount of the DC value is reduced to about a quarter as compared with the conventional method.

First, an 8-bit average value is calculated from the 64 DC values. Subtract the average value from the 64 DC values and obtain 1
Four DC vectors consisting of six values are created, and each DC vector
The vector is vector-quantized and converted to a 16-bit code. Therefore, the 64 DC values are converted into a code of a total of 72 bits of an average value (8 bits) and 16 × 4 = 64 bits.

Since the DC value is created from the image data of 16 pixels, the 64 DC values correspond to the image data of 16 × 64 = 1024 pixels. Therefore, 64 D
When the code created from the C value is converted into the code amount per pixel, (8 + 64) /64//16=0.07 bits. When encoding is performed with Y: U: V = 4: 1: 1, the code amount per pixel is 1.5 times, so that 0.
07 × 1.5 ≒ 0.1.

Therefore, since the conventional code amount obtained by the same method is 0.4 bits, the code amount is about one fourth. As a result, the compression rate of the entire code amount including the code amount of the AC vector is improved as compared with the related art.

That is, as described above, when the code amount per pixel according to International Publication No. WO 00/02393 is examined, when the AC value is 0.4 bits, the DC value is also about 0.4 bits. The total code amount was about 0.8 bits. Even if the AC value is 0.2 bits, DC
The value remained at 0.4 bits, for a total of 0.6 bits.

However, in the embodiment of the present invention,
Since the DC value can be reduced to 0.1 bits, if the code amount of the AC value is 0.4 bits, it becomes 0.5 bits in total,
When the code amount of the AC value is 0.2 bits, a total of 0.3
Become a bit. Therefore, the compression rate of the entire code amount is also improved.

Conventionally, since processing is performed in units of bits, the processing load is heavy for a general CPU.
Since the code amount can be reduced to about half, Huffman coding and decoding have been performed also in coding of DC values of image data such as JPEG.

However, in the embodiment of the present invention,
Since the Huffman coding can be omitted in the coding of the DC value, not only the compression rate is improved, but also the processing can be speeded up.

In the above embodiment, the average value is obtained for four DC vectors. However, the average value may be obtained for, for example, eight DC vectors.

The present invention is not limited to the above-described embodiment, and it goes without saying that various changes can be made within the scope of the technical idea described in the claims.

[0147]

According to the first aspect of the present invention, since a DC value can be encoded by vector quantization, the compression ratio can be improved as compared with the conventional Huffman encoding. Moreover,
In vector quantization, if the image quality is reduced, the compression ratio can be further improved. Therefore, encoding that emphasizes the compression ratio can be performed. In addition, vector quantization does not require an arithmetic processing time as compared with Huffman coding, so that the encoding process can be speeded up.

According to the second aspect of the present invention, in order to generate a DC vector, the average value in the average value separation type block coding method is the number of DC values of an integer multiple of twice or more the number of elements of the DC vector. Is required, and the average value is a DC value (DC vector element) having a small variation compared to the variation of the pixel value
The variation is even smaller than the variation of the value of, and a large numerical value can be set as a value of the integral multiple. The larger the value of the integral multiple that is set is, the larger the average value used for encoding the DC vector is.
Vector element), and the code amount can be reduced.

According to the third aspect of the present invention, the DC value is encoded by the DC vector number for specifying the vector in the DC dictionary.
The encoding is performed by using the vector number for use and the corresponding multiplier coefficient, and the compression rate can be improved by reducing the code amount.

According to the fifth aspect of the present invention, a DC vector number and a D
Since the multiplier coefficient and the average value corresponding to the C vector number are output, decoding processing such as decoding of a DC vector and decoding of a DC image can be performed using the output codes.

According to the sixth aspect of the present invention, not only the DC value but also the AC value can be encoded by the vector quantization, and the encoding in which the code amount can be reduced uniformly can be performed. In particular, since the encoding is performed by vector quantization, in which the compression rate is improved if the image quality is reduced, encoding with emphasis on the compression rate can be performed to greatly reduce the entire code amount.

According to the seventh aspect of the present invention, when the AC vector is created, the D vector created from the image data of the original image is used.
Since the AC vector is created using the DC value of the DC image decoded from the DC vector instead of the C image, the code of the AC vector in consideration of decoding the decoded image to the original image as accurately as possible is considered. If the DC value and the AC value are decoded, an image that reproduces the original image as much as possible can be obtained by decoding.

According to the eighth aspect of the invention, the same effect as that of the third aspect is obtained for the AC value, and according to the ninth aspect, the same effect as that of the fourth aspect is obtained for the AC value.
Therefore, it is possible to reduce the code amount by encoding the DC value and the AC value by vector quantization in a unified method.

According to the tenth aspect of the present invention, an AC vector number, a multiplier coefficient corresponding to the AC vector number, and a DC value are output as codes of an AC value for compressing image data. The decoding process of decoding the AC vector and decoding the original image can also be performed by using.

According to the eleventh, seventeenth and eighteenth aspects, the same effects as those of the first and sixth aspects can be obtained, and according to the twelfth, seventeenth and eighteenth aspects. The same effect as the effect of the second invention is obtained.
According to the inventions of Claims 7 and 18, the same effect as the effect of the invention of Claim 3 is obtained, and according to the inventions of Claims 14, 17, and 18, the same effect as the effect of the invention of Claim 4 is obtained, Claim 1
According to the inventions of 5, 17, and 18, the same effect as the effect of the invention of claim 5 is obtained, and according to the invention of claims 16, 17, and 18, the same effect as the effect of the invention of claim 7 is obtained. Can be

[Brief description of the drawings]

FIG. 1 is a schematic block diagram of an image data compression device according to an embodiment of the present invention.

FIG. 2 is a flowchart illustrating an outline of a process of creating a DC dictionary held by a DC dictionary holding unit.

FIG. 3 is a flowchart showing the processing of the flowchart of FIG. 2 in further detail;

FIG. 4 shows steps SS13, SS14 and SS1 of FIG.
FIG. 14 is a flowchart showing the processing of No. 5 in further detail.

FIG. 5 is a diagram for explaining a relationship among pixels, DC values, and DC vectors of image data.

FIG. 6 is a flowchart illustrating an operation of the image data compression device of FIG. 1;

FIG. 7 is a diagram showing an internal configuration of a conventional image data compression device.

FIG. 8 is a flowchart for explaining the operation of the image data compression device of FIG. 7;

FIG. 9 is a diagram illustrating an example of a Huffman tree for explaining Huffman coding.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Image data compression device 5 DC image creation part 6 DC vector quantization part 7 DC vector creation part 9 DC vector encoding part 11 DC dictionary holding part 13 DC vector decoding part 15 DC image decoding part 16 AC vector quantization part 17 AC Vector creation unit 19 AC dictionary holding unit 21 AC vector encoding unit

(54) [Title of the Invention] Image data compression apparatus, image data compression method, program for causing computer to execute image data compression method, and program for causing computer to execute image data compression method Computer-readable recording medium that records

Claims (18)

[Claims] 1. An image data compression apparatus for encoding and compressing image data, wherein the image data is divided into blocks of a plurality of pixels, and a DC image comprising a DC value representing an average of pixel values for each block is created. Image creating means; and a plurality of DCs of the DC image created by the DC image creating means.
Calculating an average of the values, subtracting the average from each of the DC values for which the average has been determined, creating a DC vector having a plurality of values obtained by the subtraction as elements, And a DC vector quantization means for vector-quantizing and encoding the image data. 2. A D value necessary for obtaining the average value.
2. The image data compression apparatus according to claim 1, wherein the number of C values is an integer multiple of at least twice the number of elements of the DC vector. 3. The DC vector quantization means includes: a DC vector creation unit that creates the DC vector; and a DC dictionary that is a set of vectors for encoding the DC vector created by the DC vector creation unit. A vector in the DC dictionary held by the DC dictionary holding unit is searched for a vector in which an error between the held DC dictionary holding unit and the DC vector created by the DC vector creating unit is within an allowable value. The image data compression apparatus according to claim 1, further comprising: a DC vector encoding unit that encodes the DC vector. 4. The DC vector encoding means according to claim 1, wherein
A vector in which an error between the DC vector created by the vector creation unit and the DC vector is within an allowable value is searched by multiplying a vector in the DC dictionary held by the DC dictionary holding unit by a multiplier coefficient. The image data compression device according to claim 3, wherein the DC vector is encoded by a DC vector number specifying each vector of a set of vectors as a dictionary and a multiplier coefficient corresponding to the DC vector number. 5. The image data according to claim 4, further comprising output means for outputting the DC vector number, a multiplier coefficient corresponding to the DC vector number, and the average value as a sign of a DC value of the image data. Compression device. 6. The method according to claim 1, wherein each of the pixel values of the image data is
An AC vector quantization means for subtracting a value, creating an AC vector having a plurality of values obtained by the subtraction as elements, and vector-quantizing and encoding the created AC vector. 6. The image data compression device according to any one of 1 to 5. 7. A DC vector decoding means for decoding a DC vector encoded by the DC vector quantization means, and a DC image decoding means for decoding a DC image from the DC vector decoded by the DC vector decoding means, 7. The image data compression device according to claim 6, wherein said AC vector quantization means uses a DC value of a DC image decoded by said DC image decoding means as said DC value. 8. The AC vector quantization means includes: an AC vector creation means for creating the AC vector; and an AC dictionary which is a set of vectors for encoding the AC vector created by the AC vector creation means. A vector in which the error between the held AC dictionary holding unit and the AC vector created by the AC vector creating unit is within an allowable value is searched from the vector in the AC dictionary held by the AC dictionary holding unit. 8. The image data compression device according to claim 6, further comprising: an AC vector encoding unit that encodes the AC vector. 9. The AC vector encoding means according to claim 1, wherein
The vector in the AC dictionary held by the AC dictionary holding unit is searched for a vector whose error from the AC vector created by the vector creating unit is within an allowable value by multiplying the vector by a multiplier coefficient. 9. The image data compression device according to claim 8, wherein the AC vector is encoded using an AC vector number for specifying each vector of a set of vectors as a dictionary and a multiplier coefficient corresponding to the AC vector number. 10. The AC vector number and the AC vector number.
10. The image data compression device according to claim 9, further comprising an output unit that outputs a multiplier coefficient corresponding to a use vector number and the DC value as a code of the image data. 11. An image data compression method for encoding and compressing image data, wherein the image data is divided into blocks of a plurality of pixels, and a DC image including a DC value representing an average of pixel values for each block is created. Step 1; and calculating an average value of a plurality of DC values of the DC image created in the first step.
A second step of subtracting the average value from the C value, creating a DC vector having a plurality of values obtained by the subtraction as elements, and vector-quantizing and encoding the created DC vector; Subtracting the DC value from each pixel value of the image data, creating an AC vector having a plurality of values obtained by the subtraction as elements, and performing vector quantization and encoding on the created AC vector. 3. A method for compressing image data, comprising: 12. The image data compression according to claim 11, wherein the number of DC values required for obtaining the average value is an integer multiple of twice or more the number of elements of the DC vector. Method. 13. A DC dictionary, which is a set of vectors for encoding the DC vector created in the first step, is prepared in advance, and the second step is performed in the first step. 2. The method according to claim 1, further comprising: searching a vector in the DC dictionary for a vector having an error between the generated DC vector and an allowable value, and encoding the DC vector.
13. The image data compression method according to 1 or 12. 14. The second step is to multiply a vector in the DC dictionary by a vector whose error between the DC vector created in the first step is within an allowable value and a multiplier coefficient by a vector coefficient in the DC dictionary. Encoding the DC vector with a DC vector number that specifies each vector of the vector set that is the DC dictionary and a multiplier coefficient corresponding to the DC vector number.
The image data compression method according to claim 13. 15. The DC vector number and the DC
15. The image data compression method according to claim 14, further comprising a fourth step of outputting a multiplier coefficient corresponding to a use vector number and the average value as a code of the image data. 16. A fifth step of decoding the DC vector encoded in the second step, and a sixth step of decoding a DC image from the DC vector decoded in the fifth step. 16. The image data compression method according to claim 11, wherein the third step includes a step of using a DC value of the DC image decoded in the sixth step as the DC value. 17. A program for causing a computer to execute the image data compression method according to claim 11. Description: 18. A computer-readable recording medium on which the program according to claim 17 is recorded.