JP2005012823A - Data compressing device and method, data expanding device and method, data compressing/expanding system and method, code book making method, and recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To reproduce data with high quality on the basis of compressed data by realizing image or voice data compression with a high compression rate. <P>SOLUTION: A correlation ratio calculation section 36 quantizes the vectors of each frame of a moving picture by using a space direction code book. A space direction code book reconstruction calculating section 39 rearranges the obtained code string to generate new vectors. The correlation ratio calculation section 36 quantizes the new vectors by using a time-axis code book. In such a way, not only vector quantization in the space direction is performed for each frame image, but also vector quantization in the time-axis direction is performed between frames along the time axis. Thus, the data compression apparatus or the like is characterized in that data compression is applied between frames so as to compress a moving picture at a high compression rate while keeping the quality of a reproduced image. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a data compression apparatus and method, a data expansion apparatus and method, a data compression / decompression system and method, a codebook creation method, and a recording medium storing a program for executing these processes, and in particular, data compression. It is suitable for use in an apparatus, system, and method using vector quantization as one of the methods.

  Conventionally, various data compression techniques have been proposed. Among them, a technique called “vector quantization” is well known as one of data compression algorithms that can perform decompression processing of compressed data very easily. This algorithm has been known for a long time in the field of signal processing, and in particular has been applied to data compression or pattern recognition of image signals and audio signals.

  In this vector quantization, several pixel patterns of a certain size (for example, a 4 × 4 pixel block) are prepared, and a unique number or the like is given to each pixel pattern (this set is referred to as a “code book”). Then, for example, blocks of the same size (for example, 4 × 4 pixels) are sequentially extracted from the image data of a two-dimensional array, and the most similar pattern is found from the code book, and the pattern number is assigned to the block. Perform data compression. In vector quantization, a data string in one block corresponds to one vector.

  On the reception side or decompression side of the compressed data encoded in this way, the original image can be reproduced simply by taking out the pattern corresponding to the number for each block from the code book. Therefore, on the decompression side, if the code book is received or held in advance, no special operation is required, and the original image can be reproduced with very simple hardware.

  In vector quantization as described above, for example, when compressing data of an image or audio signal, how to obtain a high-quality reproduced image or reproduced audio while maintaining a high compression rate, and execute vector quantization The issue is how to create a high-performance codebook that is absolutely necessary.

  Conventionally, in order to achieve a high compression rate, for example, efforts have been made to increase the size of blocks extracted from image data or to reduce the number of blocks constituting a codebook. As a codebook optimization technique, several techniques such as Kohonen's self-organizing map technique have been used.

  However, for example, when an image is compressed using a vector quantization technique, if a high compression ratio is to be realized, there arises a problem that the image quality of the reproduced image is inevitably deteriorated. Conversely, when trying to improve the quality of the reproduced image, processing other than vector quantization is performed, resulting in a problem that the compression rate does not increase.

  A vector quantization technique used for compressing image data is configured to perform vector quantization on a still image. Therefore, compression of moving images has been realized by performing vector quantization on each frame and further combining techniques other than vector quantization. Therefore, particularly in the case of moving images, there is a problem that the image quality of the reproduced image is poor even though the compression rate does not increase as a result.

  The codebook optimization technique such as Kohonen's self-organizing map technique is to optimize the codebook by performing appropriate mathematical processing using a sample image or the like. Therefore, there is a problem that the obtained code book becomes a useful code book only for the data used in the optimization.

  That is, for example, a codebook optimized using image data of a person's face is the best codebook for the image used for the optimization, but is not necessarily best for other images. Not necessarily a codebook. Therefore, for example, if the code book is used for image data of another person's face and data compression is performed, the image quality of an image reproduced from the compressed data is lowered.

  Furthermore, for images included in the classification of the same person's face as the image used for optimization, even if a relatively good image quality is obtained as a reproduced image, for images of different classifications such as landscapes and characters In many cases, the image quality deteriorates. That is, since the code book pattern is completely different depending on the image, there is a problem that the code book becomes less versatile.

  In addition, in the vector quantization technique using a code book, a process of finding a code vector having a pattern most similar to an input vector from the code book, and a process of extracting a code vector corresponding to a compressed code from the code book and displaying the code vector However, there is a problem that it takes a lot of operations and takes time.

The present invention has been made to solve such a problem. For example, when data compression of an image or an audio signal is performed by vector quantization, the data compression is realized at a high compression rate. It is an object to enable reproduction of high quality data based on the obtained compressed data.
Another object of the present invention is to realize a highly versatile code book that can handle various images.
Another object of the present invention is to improve the processing speed at the time of compression by vector quantization and image reproduction by expansion.

  In the first aspect of the present invention, a data string having at least one or more data is blocked into a vector, and a code vector similar to a vector extracted from a compression target is searched from a code book prepared in advance. In the data compression apparatus that outputs a code corresponding to the code book, the code book is a code book optimized by repeating a predetermined number of times of update processing using a predetermined learning algorithm, and is a compression that changes along a time axis. For each target data at each time, a plurality of blocks are extracted from the data, code vectors similar to those vectors are searched from the first code book, and a code string corresponding to each code vector is found for each time. First vector quantization means for outputting to the first vector quantization means and the first vector quantization means The code string output at each time is rearranged into a plurality of new vectors, and for each of the plurality of new vectors, a code vector similar to each of the vectors is searched from the second code book. And a second vector quantization means for outputting a code string corresponding to them.

  In the second invention, a data string having at least one or more data is blocked into a vector, and a code vector similar to a vector extracted from a compression target is searched from a code book prepared in advance. In the data compression device that outputs a corresponding code, the code book is a code book that is optimized by repeating an update process using a predetermined learning algorithm a predetermined number of times, and is a compression target that changes along a time axis. For each data at each time, a plurality of blocks are extracted from the data, code vectors similar to those vectors are searched from the first code book, and a code string corresponding to each is extracted for each time. First vector quantization means for outputting, and the first vector quantization means Among the code strings output at each time, the code string of data at a certain time is used as a reference code string, and between the codes at the corresponding addresses between the reference code string and the code string of data at another time The result of the difference is rearranged into a plurality of new vectors, and for each of the plurality of new vectors, code vectors similar to those vectors are respectively searched from the second codebook, and corresponding to them. And a second vector quantization means for outputting a code string.

  According to a third aspect of the present invention, a data string having at least one or more data is blocked into a vector, and a code vector similar to a vector extracted from a compression target is searched from a code book prepared in advance. In the data compression device that outputs a corresponding code, the code book is a code book that is optimized by repeating an update process using a predetermined learning algorithm a predetermined number of times, and is a compression target that changes along a time axis. For each data at each time, a plurality of blocks are extracted from the data, code vectors similar to those vectors are searched from the first code book, and a code string corresponding to each is extracted for each time. First vector quantization means for outputting, and the first vector quantization means From among the code strings output at each time, the result of taking the difference between the codes of the corresponding addresses between the data adjacent in the time axis direction is rearranged into a plurality of new vectors, And a second vector quantization means for searching each new vector for code vectors similar to those vectors from the second codebook and outputting a code string corresponding to the code vectors. To do.

According to a fourth aspect of the present invention, there is provided a time axis shift means for generating the plurality of new vectors by shifting the elements in the new vector at least between adjacent vectors so as to be shifted in the time axis direction. It is characterized by that.
In addition, when generating a plurality of new vectors on the decompression side that performs processing opposite to that of the data compression apparatus according to the first to third inventions, at least adjacent elements in the new vector are represented on a time axis. It is characterized by comprising time axis shift means for shifting the arrangement in the direction.

  According to a fifth aspect of the present invention, a data string having at least one or more data is blocked into a vector, and a code vector similar to a vector extracted from a compression target is searched from a code book prepared in advance. In the data compression device that outputs a corresponding code, the code book is a code book that is optimized by repeating an update process using a predetermined learning algorithm a predetermined number of times, and at least one or more codebooks from the compression target A separation means for extracting feature quantities and separating them into feature pattern data and basic pattern data obtained by removing the feature quantities from the compression target is provided, and a vector is independently provided for each of the separated feature quantity data and the basic pattern data. It is characterized by performing quantization.

According to a sixth aspect of the present invention, a data string having at least one or more data is blocked into a vector, and a code vector similar to a vector extracted from a compression target is searched from a code book prepared in advance. In the data compression device that outputs a corresponding code, the code book is a code book that is optimized by repeating an update process using a predetermined learning algorithm a predetermined number of times, and is a compression target that changes along a time axis. A vector that extracts a plurality of blocks from the data at each time, searches for code vectors similar to those vectors from the code book, and outputs a corresponding code string at each time Output at each time by the quantization means and the vector quantization means. Output control means for taking a correlation between the codes corresponding to addresses adjacent to each other in the time axis direction and outputting a code only for an address where the correlation is smaller than a predetermined value. It is characterized by that.
In another aspect of the sixth aspect of the present invention, the feature quantity of the code vector found at each time by the vector quantization means is correlated with the feature quantity between corresponding addresses of data adjacent in the time axis direction. And output control means for outputting a code only for an address having a correlation smaller than a predetermined value.

  According to a seventh aspect of the present invention, a data string having at least one or more data is converted into a block to obtain a vector, and a code vector similar to a vector extracted from a color image to be compressed is searched from a code book prepared in advance. In the data compression apparatus that outputs a code corresponding to the code book, the code book is a code book that is optimized by repeating a predetermined number of update processes using a predetermined learning algorithm, and a code book for luminance signals And a code book for color signals, and more code vectors are assigned to the code book for luminance signals than the code book for color signals.

An eighth invention is a method for creating a codebook which is composed of a set of vectors, which is a data string having at least one data, and is used in vector quantization, and includes a sample data and an initial codebook. The code book is optimized by repeatedly performing vector quantization processing and updating the contents of the specified range on the virtual two-dimensional plane of the code book according to a predetermined update coefficient for each processing. In the method of creating a code book to be converted, a code book having a pattern in which data values continuously change from a minimum value to a maximum value of possible values is used as the initial code book.
In addition, the range to be updated by the one-time process is applied one-dimensionally on the virtual two-dimensional plane, and the range is decreased as the number of updates increases.
The initial value of the update coefficient is set to any value within a range of 0.3 to 1, and the value of the update coefficient is decreased as the number of updates increases.

  A ninth invention is a method for creating a codebook which is composed of a set of vectors, which is a data string having at least one or more data, and is used in vector quantization, and includes a sample data and an initial codebook. The code book is optimized by repeatedly performing vector quantization processing and updating the contents of the specified range on the virtual two-dimensional plane of the code book according to a predetermined update coefficient for each processing. In the method of creating a code book to be converted, a code book is optimized independently for a plurality of sample data, and a plurality of obtained code books are synthesized to create a new code book. .

  According to a tenth aspect of the present invention, feature quantity storage means for storing a feature quantity obtained in advance for each code vector in a codebook, feature quantity calculation means for obtaining a feature quantity of a vector extracted from the compression target, and the feature Based on the feature quantity of each code vector stored in the quantity storage means and the feature quantity of the compression target vector obtained by the feature quantity calculation means, the similarity between each code vector and the compression target vector And a calculation omitting means for determining whether or not to omit the calculation for obtaining.

  An eleventh aspect of the present invention is a data compression / decompression system comprising a codebook server that holds at least one codebook, a data compression system, and a data decompression system, wherein the codebook is a predetermined codebook The code book is optimized by repeating the update process using a learning algorithm a predetermined number of times, and the code book server holds any code held in accordance with a request from the data compression system and the data decompression system. It is characterized by supplying a book.

  In a twelfth aspect based on the sixth aspect, the output control means includes the vector quantization means for all or a part of the addresses regardless of whether the correlation is smaller than a predetermined value at a certain time. The output code string is output.

  As described above in detail, according to the first to third inventions, when vector quantization is performed on a compression target that changes along the time axis, a vector in the spatial direction is applied to one data string within the same time. Since not only quantization but also vector quantization in the time axis direction between a plurality of data strings along the time axis, for example, not only data compression within each frame of a moving image, Data compression can also be performed between frames. As a result, the moving image can be compressed at a higher compression rate while maintaining the quality of the reproduced image.

  According to the fourth aspect of the present invention, when generating a plurality of new vectors to perform vector quantization in the time axis direction at the time of compression, time is required between at least adjacent vectors in the new vector. When arranging a plurality of new vectors when shifting in the axial direction or generating a plurality of new vectors at the time of decompression and processing opposite to those at the time of compression, at least adjacent elements in the new vectors Since, for example, a moving image is used as a compression target, the quantization error that is conspicuous each time an image is reproduced at the time of decompression may be inconspicuous. Thus, it is possible to obtain a high-quality reproduced image while maintaining a high compression rate of the moving image.

  According to the fifth invention, at least one feature quantity is extracted from the compression target and separated into feature quantity data and the remaining basic pattern data, and the separated feature quantity data and basic pattern data are separated. Since the vector quantization is performed independently, the codebook used for vector quantization of each data can be simplified compared to the codebook for the compression target itself in which various features are intertwined. The number of necessary patterns can also be reduced. As a result, when searching for a pattern having a high degree of similarity in the code book by vector quantization, it becomes easier to narrow down to a more appropriate pattern, and highly accurate vector quantization can be performed. Therefore, for example, when image data is used as a compression target, a higher quality reproduced image can be obtained while maintaining a high compression rate.

  According to the sixth aspect of the invention, vector quantization is performed on the data at each time to be compressed that changes along the time axis, whereby the code string obtained at each time or the code string is obtained. For the feature quantity of the corresponding code vector, the code is output only for the portion where the correlation in the time axis direction is small. For example, when moving images are used as compression targets, only the data where the data value has changed significantly between frames Can be transferred, and data with little change can be prevented from being transferred. As a result, in addition to the result of vector quantization in the spatial direction within each frame, data can be compressed between frames, and moving images can be compressed and transferred at a high compression rate. .

  Also, according to the seventh aspect, when vector quantization of a color image is performed, more code vectors are assigned to a code book for luminance signals than to a code book for color signals. A code vector having a higher similarity with respect to a luminance signal having a higher importance can be obtained by vector quantization. Therefore, by not changing the overall code length, it is possible to improve the quality of the reproduced image while maintaining the compression rate.

  According to the eighth invention, when creating a code book, a code book having a pattern in which data values continuously change from a minimum value to a maximum value of possible values is used as an initial code book. The range to be updated by the process is applied one-dimensionally on a virtual two-dimensional plane, and the range is decreased as the number of updates increases, and the initial value of the update coefficient is any value within the range of 0.3 to 1 Since the update coefficient value is decreased as the number of updates increases, the code book can be optimized more efficiently and a highly accurate code book can be obtained. Thereby, the image quality of a reproduced image can be improved by using the generated code book in, for example, an image data compression / decompression system.

  According to the ninth invention, when a code book is created, the code book is optimized independently for a plurality of sample data, and a plurality of obtained code books are synthesized to create a new code book. Since it was made to do, the code book which has tolerance with respect to various images can be created. Thereby, a general-purpose code book can be obtained without increasing the size of the device for storing the code book.

  According to the tenth invention, the feature amount of the vector extracted from the compression target is compared with the feature amount of each code vector in the code book, and each code vector is determined based on the comparison result. Since it is determined whether or not to omit the calculation for calculating the similarity with the vector to be compressed, the calculation of the similarity is appropriately omitted by calculating the feature amount having a smaller calculation amount than the calculation of the similarity. When performing an operation to search for a code vector similar to the vector to be compressed, the amount of calculation required for the search can be reduced, and the vector quantization operation as a whole can be performed at high speed. .

  Further, according to the eleventh invention, since the code book which is always necessary when performing compression / decompression by vector quantization is prepared separately from the compression / decompression system, each compression / decompression system uses a code book. There is no need to hold the data, it is only necessary to transfer actual data, and data transfer can be performed using a medium having a smaller capacity. As a result, a very simple data transfer system can be realized.

  According to the twelfth aspect, with respect to the ninth aspect, all or a part of the addresses output from the vector quantizing means are output regardless of whether the correlation is smaller than a predetermined value at a certain time. Since the code string is output as update data, even if the state where the correlation is smaller than the predetermined value and the data of the previous frame is used as it is continues, the state can be interrupted by transmission of the update data, It is possible to suppress the deterioration of the image quality with the progress of.

(First embodiment)
In the first embodiment, when vector quantization is performed on a still image, the vector quantization is normally performed on a square block such as 4 × 4 pixels. Thus, vector quantization is performed on a block consisting of one line. Further, at the time of vector quantization or when the compressed image is reproduced, a process of shifting the block for each line in the horizontal direction of the image is performed.

  FIGS. 1 and 2 are functional block diagrams showing configurations of the data compression apparatus and the decompression apparatus according to the first embodiment, respectively. FIGS. 3 and 4 are data compression apparatuses shown in FIGS. 1 and 2, respectively. 5 is a flowchart showing the operation of the data decompression apparatus. 5 and 6 are explanatory diagrams for explaining the operation on the expansion side according to the present embodiment.

  First, FIG. 5 and FIG. 6 will be described. In FIG. 5, when a line block along a scanning line on an image is used as vector data as vector data (line block method of the present embodiment), a conventional square block is used as vector data (macroblock method). It explains in comparison with. FIG. 5A shows a conventional macroblock system, and FIG. 5B shows a line block system of this embodiment.

  As shown in FIG. 5A, the code book memory stores, for example, a large number of code vectors in the order of addresses, with 16 data strings constituting a macroblock of 4 × 4 pixels as one code vector. . When a playback image is generated using this code vector and displayed on a display or the like, address conversion is performed on the display memory so that 16 data strings are displayed at the correct pixel position of 4 × 4. Had to do.

  On the other hand, in the line block system of this embodiment shown in FIG. 5B, the vector quantization is performed by extracting a block on one line of 16 × 1 pixels. Since the data arrangement in this line block matches the scanning direction of the display, the address conversion operation required in the case of a square macroblock can be avoided. The code book in this case can also be easily obtained by using Kohonen's self-organizing map technique.

  FIG. 7 is a diagram showing the results of measuring the time required for each process when compressing, expanding, and displaying one image of 640 × 480 pixels. Referring to FIG. 7, in the conventional macroblock system, it takes about 50 msec for vector quantization (VQ), about 160 msec for decoding, and about 60 msec for display. The reason for this long time for decoding is that, as described above, address conversion is required when the code vector specified by the code number is extracted from the code book memory and displayed on the display.

In such a case, a dedicated chip for address conversion is generally prepared to increase the speed, but in this embodiment, the above-described line block method is used as a means for speeding up more easily. Has proposed.
FIG. 7 also shows each processing time when the line block method is adopted. Although the time required for vector quantization and display scanning is not different from that in the case of the macroblock method, it can be seen that the decoding time can be greatly reduced from 160 msec.

  Also, in both cases of the macro block method and the line block method (particularly when the line block method is used), when a playback image is displayed by orderly pasting pattern images of code vectors in order, both ends of each line of the playback image are displayed. The big discontinuity of the pattern of this will be conspicuous. Therefore, as shown in FIG. 6, in the present embodiment, the code vector pattern paste position is shifted little by little for each line (hereinafter referred to as a slide block method).

  As described above, if the code vector blocks are shifted and fitted in the horizontal direction of the image at least between adjacent lines, the continuity of each block boundary is cut and the block boundary (quantization error) becomes inconspicuous. As a result, a fluffy portion can be seen at both ends of the reproduced image, but a generally good image can be obtained. Note that the disturbance at both ends of the image can be easily removed by applying a mask when displaying.

  In the present embodiment, the process of shifting the block in the horizontal direction of the image between adjacent lines is performed not only when the image after vector quantization is reproduced as described in FIG. 6, but also when the vector quantization is performed. . That is, the blocks are cut out properly in order from the upper left of the image, but the blocks are cut out in the horizontal direction of the image and vector quantization is performed.

  Next, the configuration and operation according to the first embodiment described above will be described mainly based on FIGS. In FIG. 1, reference numeral 1 denotes an image input unit for inputting original image data to be compressed. The input image data is temporarily stored in the image memory 2, read out by the address control from the reading unit 3, and output.

  Reference numeral 4 denotes a blocking unit, which divides the image data read from the image memory 2 into, for example, 16 × 1 pixel unit line blocks. At that time, at least the blocks are cut out while shifting in the horizontal direction of the image between adjacent lines. If the position of the block is shifted from the adjacent horizontal line, the shifting direction and amount can be arbitrarily set. Although the blocking unit 4 is provided here, when the reading unit 3 reads image data, if the reading is performed while shifting the line block as described above, the blocking unit 4 is not always necessary. is not.

  A code book storage unit 5 stores in advance a plurality of pixel patterns having the same size as a line block extracted from an original image (16 × 1 pixel block) as a code vector. A vector quantization unit 6 compares the data of one line block extracted from the original image by the blocking unit 4 with the data of a plurality of code vectors stored in the codebook storage unit 5. As a result, a pattern most similar to the block data of the original image is found from the code book, and the code number of the pattern is output. This operation is performed for all line blocks extracted from the original image.

  In FIG. 2, reference numeral 7 denotes a code number input unit, which inputs a code number string for each block specified on the compression side. The input code number string is temporarily stored in the code memory 8 and is read and output by address control or the like from the reading unit 9. Reference numeral 10 denotes a code book storage unit, which stores in advance a plurality of pixel patterns of blocks made up of, for example, 16 × 1 pixels as code vectors. Prior to the decompression process, the code book may be transmitted and stored on the compression side, or the same code book may be stored from the beginning.

  Reference numeral 11 denotes a reproduction image generation unit, which performs a process of reading out and applying the corresponding code vector pattern images from the code book storage unit 10 based on the code number sequence read from the code memory 8. Reference numeral 12 denotes a block shift unit, which performs a process of shifting each block constituting the reproduction image output from the reproduction image generation unit 11 in the horizontal direction of the image. At this time, for example, if the position of the block is shifted between at least adjacent horizontal lines, the shifting direction and the amount can be arbitrarily set. The reproduced image subjected to such a shift process is temporarily stored in the image memory 13 and then sent to a display device (not shown) for display.

  In FIG. 3 showing the operation on the compression side shown in FIG. 1, first, initialization processing is performed in step S1. Here, a block counter (not shown in FIG. 1) for counting processed blocks, an image buffer (image memory 2 in FIG. 1), and a code buffer for storing code numbers determined by vector quantization A process of clearing each (not shown in FIG. 1) is performed.

  Next, in step S2, the original image is input and stored in the image buffer. In step S3, a block having a size of x × y pixels (x and y are arbitrary values, but x = 16 and y = 1 in the above example) is read from the original image stored in the image buffer. . At the time of reading, the position of the block is adjusted so that the position of the block shifts at least between adjacent horizontal lines.

  In step S4, vector quantization processing is performed on the read block data, and the code having the highest correlation with the read block data from the code book stored in the code book storage unit 5 is obtained. Find the vector and determine the corresponding code number (Winner code). Next, in step S5, the obtained code number is stored in the code buffer. The code number stored here is then output to the outside. This code buffer is not always necessary.

  Next, the process proceeds to step S6, and it is determined whether or not the above-described processing has been completed for all the blocks in the image. If not completed, the value of the block counter is increased by 1 in step S7, and the process returns to step S3 to repeat the same processing. In this way, by determining and outputting code numbers for all blocks in the image, the original image is compressed.

  In FIG. 4 showing the operation on the expansion side shown in FIG. 2, the initialization process is first performed in step S8. Here, a block counter (not shown in FIG. 2) for counting processed blocks, a reproduction image buffer (image memory 13 in FIG. 2), and a code buffer (FIG. 2) for storing an input code number. The code memory 8) of 2 is cleared.

  Next, in step S9, the code number string generated on the compression side is input and stored in the code buffer. In step S10, decoding is performed based on one code number indicated by the current block counter in the stored code number string. That is, the code vector (pattern image) corresponding to the code number is extracted from the code book stored in the code book storage unit 10 and stored in the reproduction image buffer. At that time, the extracted block of x × y pixels is stored so that the position is shifted at least between adjacent horizontal lines.

  Here, in the present embodiment, the block shifting method is set to be opposite to each other in the step S3 on the compression side and the step S10 on the expansion side. That is, if the block line is shifted by i pixels in the right direction in step S3, the block line is shifted in the left direction by i pixels in step S10. By doing so, since there is no shift for each block line in the reproduced image to be reproduced, not only the quantization error can be made inconspicuous but also the quality of the reproduced image can be further improved. .

  Next, the process proceeds to step S11, and it is determined whether or not the above processing has been completed for all blocks in the image. If not completed, the value of the block counter is increased by 1 in step S12, and the process returns to step S10 to repeat the same processing. In this way, by extracting the pattern image from the code book for all the blocks in the image and storing it in the reproduction image buffer, the reproduction image stored here is given to the display or the like and displayed.

  In the above example, the block is shifted on both the compression side and the expansion side. However, the block may be shifted only on the compression side or only on the expansion side. In this case, although the reproduced image may include a shift for each block line, the quantization error itself can be made inconspicuous, and the quality of the reproduced image can be improved as compared with the conventional example.

Further, here, the data compression of the still image is described, but in the case of a moving image, vector quantization is performed on each frame. Therefore, the data compression of the moving image can be similarly applied.
In the present embodiment, a system is realized in which a moving image is compressed by vector quantization using the vector quantization PCI bus board 22 shown in FIG. 8, and the moving image is reproduced based on the result of the compression. The details will be described below.

  On the vector quantization board 22 shown in FIG. 8, for example, eight chips capable of mounting 256 code vectors of 8 bits and 16 dimensions (4 × 4 pixels) are mounted, and a total of 2048 code vectors are used. Thus, vector quantization can be performed. These chips are connected to a computer (not shown) through a PCI bus interface configured by, for example, an FPGA (Field Programmable Gate Array).

  When performing vector quantization, first, all code vectors constituting a code book prepared in advance are transferred from the memory 21 onto the chip of the vector quantization board 22. Next, the moving image data stored in advance in the memory 21 is sequentially sent to the vector quantization board 22 to execute vector quantization. The code number obtained as a result of vector quantization is temporarily stored in the memory 21. Based on this code number, the corresponding code vector is read from the memory 21 and decoded. The decoded reproduced image data is sent to the display 23 of the computer and displayed as an image.

  Here, the number of code vectors constituting the code book is 2048 and the number of data constituting the code vector is 16, but this is one specific example, and the numerical values described here It goes without saying that it is not limited to this. In addition, the image size used in the system configured here and the block size when performing vector quantization are also specific examples, and it goes without saying that they are not limited to the numerical values described here.

  As described above, in the first embodiment, a block which is a unit of vector quantization is a line block such as 16 × 1 pixel. Since this matches the display scanning direction, there is no need for address conversion processing during playback, and the image reproduction speed increases (compared to the case where vector quantization is performed with a macroblock of 4 × 4 pixels). About 3.5 times faster reproduction).

  For example, when data compression is performed on a moving image having one frame of 640 × 480 pixels by the macro block method, the frame rate is about 4 frames / second, but according to the line block method of the present embodiment, 14 frames per second. A vector quantization operation of about the frame can be realized.

  The processing time of the display portion can be shortened by improving the display program. In the current system, about 60% of the time required for vector quantization processing is occupied by address conversion and transfer time for transferring the input vector to the vector quantization board. Is only a few working. Therefore, it is possible to further reduce the processing speed by expanding the transfer bus width and increasing the processing capacity of the CPU. By performing these improvements, a vector quantization operation of 30 frames per second can be realized.

  Also, in this embodiment, each block is shifted in the horizontal direction of the image and cut out to perform vector quantization, or the image is reproduced while shifting each block. Vertical lines (quantization errors) due to conspicuous block boundaries at both ends are dispersed and become inconspicuous, and image quality can be improved.

  In the above embodiment, the line block method and the slide block method are used together, but only one of them may be used. When only the line block method is employed, at least the image reproduction speed can be increased, and when only the slide block method is employed, at least the quantization error can be made inconspicuous.

  In the above embodiment, an example is shown in which blocks are shifted with respect to the spatial direction within one image. However, in the case of compressing data of a moving image having a plurality of frames, the time axis direction between the frames is shown. On the other hand, the elements in the block may be shifted. Even in this case, the quantization error in the reproduced image can be made inconspicuous.

(Second Embodiment)
When compressing a moving image, it is difficult to obtain a high compression rate only by performing vector quantization on each frame (still image) constituting the moving image. Therefore, in the second embodiment, not only vector quantization within a frame is performed on each still image of a plurality of frames, but also data between the same addresses of each still image of a plurality of frames is formed as a vector, However, data compression is performed by applying vector quantization.

  FIG. 9 and FIG. 10 are functional block diagrams showing a configuration example of the data compression / decompression system according to the second embodiment, and FIG. 11 is a data flow showing the data flow at the time of compression according to the second embodiment. FIG. 9 shows the configuration on the compression side, and FIG. 10 shows the configuration on the expansion side. In addition, the flow of data at the time of decompression is basically the reverse of the flow at the time of compression, and can be assumed from FIG. 11, so the data flow diagram at the time of decompression is not particularly shown here.

  Here, as shown in FIG. 11, a case will be described in which compression / expansion is performed on a moving image in which one frame is composed of 640 × 480 pixels and the luminance value per pixel is represented by 8 bits. Here, for convenience of explanation, it is assumed that compression / expansion is performed on an image of 16 frames.

  In FIG. 9, reference numeral 31 denotes an image input unit, which sequentially inputs original image data to be compressed, in this case, moving image data, for each frame. The input image data is temporarily stored in the image memory 32 for each frame, and is read and output by address control or the like from the reading unit 33.

  Reference numeral 34 denotes a blocking unit, which divides the image data read from the image memory 32 into blocks of, for example, 4 × 4 pixels. Here, the blocking unit 34 is provided. However, when the reading unit 33 reads image data, the blocking unit 34 is not necessarily required if reading is performed in units of blocks.

  A codebook storage unit 35 for space direction is a codebook (for performing vector quantization in one frame space) for performing a vector quantization within one frame on a single still image (frame). Since this is a code book, this is hereinafter referred to as a “space direction code book”) and stored in advance as a database.

  Here, this spatial direction codebook is a combination of a plurality of code vectors, with a data string of 16 data as one code vector. Here, a codebook for spatial direction is configured by combining 2048 vectors.

  Reference numeral 36 denotes a correlation degree calculation unit. When a mode for performing spatial direction vector quantization is specified by the calculation mode specification unit 37, the spatial direction codebook storage unit 35 stores each frame sequentially input. Using the stored spatial direction code book, the correlation, that is, the similarity between a plurality of code vectors in the code book and the input vector from the original image is calculated.

  Here, the similarity means, for example, a value obtained by quantifying how similar two are obtained by inputting two vector data into a certain function. A representative example of the above function is a function for obtaining a Manhattan distance (difference absolute value distance) or Euclidean distance between two input vector data. For example, for the Manhattan distance, the absolute value of each difference between the data of the individual elements constituting the input vector and the data of the individual elements constituting the code vector is obtained, and the obtained absolute values of the respective differences are added. Can be obtained.

In order to find the code vector most similar to the input vector, for example, the Manhattan distance as described above is obtained for each code vector constituting the spatial direction code book, and the one having the smallest distance is found.
The function for obtaining the similarity is not limited to the above function, and any function can be used according to the purpose. For example, a function in which a coefficient (weight) is added to one or both vector data, a function for calculating feature amount data (for example, the sum of elements of vector data) in a block, or the like may be used.

  A code determination unit 38 outputs a code number corresponding to the code vector having the highest similarity as a compressed code (Winner code) of the block based on the calculation result of the similarity by the correlation calculation unit 36. The correlation calculation unit 36 and the code determination unit 38 constitute a vector quantization unit. As a vector quantization method here, a conventional method may be used, or the method described in the first embodiment may be used.

  Hereinafter, this spatial direction vector quantization operation will be described in detail with reference to FIG. First, vector quantization is performed on the first frame image using the spatial direction codebook. Here, the block area obtained by dividing the image of the first frame by 4 pixels in the horizontal direction and 4 pixels in the vertical direction is defined as an input vector, and one input vector and a spatial direction codebook are configured. Similarity with each code vector is obtained, and the code number of the code vector most similar to the input vector is specified.

  By performing this operation on all input vectors in the frame, each block area obtained by dividing the image of the first frame into 4 × 4 pixels is assigned a code number corresponding to the code vector in the spatial direction codebook. Replace with. In FIG. 11, the numerals 127, 198,... Indicate the code numbers replaced for the respective blocks extracted in order from the upper left to the lower right of the image.

  Next, the same processing as that of the first frame image is performed on a total of 15 images from the second frame image to the 16th frame image, and each frame image has 4 in each frame. Each block area divided into 4 pixels is replaced with a code number corresponding to the code vector in the spatial direction codebook.

  Returning to FIG. 9, reference numeral 39 denotes a spatial direction code rearrangement operation unit, which is a block indicated by the same address for each frame image represented by the code number of the spatial direction code book by the spatial direction vector quantization means. By rearranging the code numbers in the order of frame images, the code numbers of the same address blocks distributed in the time axis direction are collectively expressed as one data string.

  That is, when spatial direction vector quantization is performed on an image for 16 frames as shown in FIG. 11, the code number of the block at the specified address in the image of the first frame is The code number of the block having the same address as the image of the first frame in the second frame image is placed at the position of the first address of the newly created data string. In this way, rearrangement is performed on the code numbers of the blocks at the same address for 16 frames, and the code numbers of the blocks at the same address in the image of the last 16th frame are 15 in the newly created data string. It will be placed at the address.

  A set including the 16 code number strings newly generated in this way is set as a new vector. Hereinafter, this is referred to as a “time axis vector”. In the case of the example shown in FIG. 11, for example, the code numbers of the blocks specified by the head addresses of the first to sixteenth frames are rearranged (127, 287, 58,..., 1483, 876). ) Corresponds to one time axis vector.

  As in this example, if one frame is an image of 640 × 480 pixels and one block area is 4 × 4 pixels, 160 × 120 = 19200 blocks are generated from one frame image. The Therefore, when the rearrangement process as described above is performed on the entire range of the image subjected to spatial direction vector quantization, 19200 time-axis vectors are newly generated.

  Returning to FIG. 9 again, reference numeral 40 denotes a time-axis codebook storage unit, which is a codebook for performing vector quantization on the time-axis vector created as described above (a plurality of frames given over time). Since this is a code book for performing vector quantization on data in between, this is hereinafter referred to as a “time axis code book”) and stored in advance as a database. This time-axis codebook is also a combination of 2048 code vectors, with a data string of 16 data as one code vector.

  When the spatial direction code rearrangement calculation unit 39 rearranges the entire range of the image, a time-axis vector quantization start signal is output and supplied to the calculation mode designating unit 37. In response to this, the calculation mode designating unit 37 switches to a mode for performing time axis vector quantization. When the mode for performing time axis vector quantization is designated by the operation mode designating unit 37, the correlation degree computing unit 36, for each time axis vector created by the spatial direction code rearrangement computing unit 39, Vector quantization is performed using the time-axis codebook stored in the time-axis codebook storage unit 40.

  Hereinafter, the operation of the time axis vector quantization will be described in detail with reference to FIG. That is, of the 19,200 time axis vectors generated previously, the similarity between one time axis vector and each code vector constituting the time axis codebook is calculated, and the similarity is the highest. The address (code number) in the code book for the time axis of the large code vector is specified. The calculation of the similarity is as described above.

  By performing this operation on all time axis vectors, a plurality of time axis vectors represented by the code number sequence after spatial direction vector quantization correspond to each code vector in the time axis codebook. Replace with code number column. In the example of FIG. 11, for example, the time axis vector (127, 287, 58,..., 1483, 876) described above is replaced with a time axis code number “10”. At this stage, an image of 640 × 480 pixels for 16 frames is represented by, for example, a data string (10,984,...) Composed of 19200 data expressed in 11 bits, and this is supplied to the decompression side. The

  Therefore, when transferring a moving image of 16 frames from the compression side to the decompression side, the space direction codebook and the time axis codebook are sent to the decompression side in advance (the same one from the beginning on the decompression side). Then, a list expressed by a code number sequence of the time axis codebook obtained by the above-described series of processes may be sent. As a result, it is possible to compress an image for 16 frames having an information amount of 39321600 bits, while maintaining the image quality, to a code number string of a time axis codebook having an information amount of 422400 bits, which is approximately 0.01 times. it can.

  On the decompression side shown in FIG. 10, the original image for 16 frames can be reproduced based on the data transferred from the compression side in this way. First, the time axis code input unit 41 inputs the code number string of the time axis code book transferred from the compression side. The input time axis code number string is temporarily stored in the time axis code memory 42, and is read and output by address control or the like from the reading unit 43.

  Then, the space direction code reproduction unit 44 uses the time axis code number sequence read from the time axis code memory 42 to transfer the code vector corresponding to each time axis code number in advance from the compression side, for example. It is taken out from the time axis code book stored in the axis code book storage unit 45. Then, the code numbers in the spatial direction are reproduced by sequentially arranging the extracted code vectors.

  Here, the time axis code number sequence read from the time axis code memory 42 includes 19200 time axis code numbers, and one code of the time axis code book corresponding to these time axis code numbers. The vector is composed of 16 pieces of data. For this reason, the code vector data arranged in order by the processing of the spatial direction code reproduction unit 44 is composed of 307200 data (spatial direction code numbers) in total.

  The spatial direction code number sequence reproduced in this way is temporarily stored in the spatial direction code memory 46, and is read and output by address control or the like from the reading unit 47. Next, the reproduction image generation unit 48 extracts data every 16 pieces from the 307200 data strings arranged as described above, generates one data string from 19200 data, and uses this as a vector. Hereinafter, this is referred to as a “space direction vector” with respect to the time axis vector described above.

  By performing this operation over the entire space of 307200 data strings, the spatial direction code number string read from the spatial direction code memory 46 is divided into 16 spatial direction vectors. The 16 spatial direction vectors generated here are sets of spatial direction code number sequences for reproducing images from the first frame to the 16th frame, respectively.

  For each of the 16 spatial direction vector strings, the reproduced image generation unit 48 further selects, for example, a code vector corresponding to each spatial direction code number (19200) constituting the spatial direction vector. From the code book for space direction stored in advance in the space direction code book storage unit 49. Then, the original image is reproduced by sequentially fitting the data of the corresponding code vector (pattern image) into each block position.

  By performing such a process every time a spatial direction vector sequence divided into 16 pieces is obtained, the original image of 16 frames can be reproduced. If the reading unit 47 performs an operation for extracting data from 307200 spatial direction code number strings every 16th, the reproduction image generation unit 48 reproduces each frame every time 19200 data is read out. Images can be generated sequentially.

  The images reproduced in this way are temporarily stored in the image memory 50 and then sent to a display device (not shown) for display or sent to a storage device (not shown) for storage.

  In the second embodiment, the number of data constituting one vector is 16, and the space direction code book and the time axis code book are each composed of 2048 code vectors. However, this is an example given for the sake of clarity in describing the present embodiment, and it goes without saying that it should not be limited to the numerical values described here.

  Also, here, compression of 16 frames of images by vector quantization has been described, but the number of frames to be compressed is not limited to 16 frames, and the required number of frames is compressed by vector quantization. Needless to say, you can. Although an embodiment using a moving image is shown here, the present invention is not limited to a moving image and can be similarly applied to an audio signal.

  In the second embodiment, one correlation degree calculation unit 36 is reused for space direction vector quantization and time axis vector quantization, but two are provided for each vector quantization. Also good.

(Third embodiment)
FIG. 12 is a data flow diagram showing a data flow during compression according to the third embodiment. The configuration of the data compression / decompression system according to the third embodiment is substantially the same as that of the second embodiment, and is not shown here. However, since the processing contents in the spatial direction code rearrangement calculation unit 39 and the reproduction image generation unit 48 are slightly different in the third embodiment, this will be described in detail.

  In this embodiment, as shown in FIG. 12, compression is performed on images for 17 frames. Therefore, in the correlation degree calculation unit 36, when the calculation of the spatial direction vector quantization is specified by the calculation mode specifying unit 37, a total of 17 images from the first frame image to the 17th frame image are obtained. A code corresponding to the code vector in the spatial direction codebook is obtained by performing spatial direction vector quantization processing on each image, thereby dividing each block region into 4 × 4 pixels for each frame image. Replace with a number.

  The spatial direction code rearrangement calculation unit 39 performs the first frame (for the code number of the block indicated by the same address in the image of each frame represented by the code number of the spatial direction code book by the correlation calculation unit 36. The difference between the code number of each frame) and the code number of each frame is calculated, and the results are rearranged in the frame image order. Note that the difference from the reference frame may be calculated for each address after the rearrangement.

  That is, as shown in FIG. 12, first, a difference is taken between the code numbers of the same addresses in the first frame and the second frame, and then the same addresses are assigned in the first frame and the third frame. Take the difference between the code numbers. Similarly, the difference is taken between the code numbers of the same addresses in the first frame and the fourth to 17th frames. Then, the difference results of the blocks indicated by the same address are rearranged in the order of frame images, so that the difference code numbers of the same address blocks distributed in the time axis direction are collectively expressed as one data string.

  That is, when spatial direction vector quantization is performed on images for 17 frames as shown in FIG. 12, the difference between the code number of the first frame image and the code number of the second frame image is The difference between the code number of the image of the first frame and the code number of the image of the third frame is placed at the position of the first address of the newly created data string. .

  In this way, rearrangement is performed on the difference between the code number of the first frame and the code numbers of the remaining 16 frames, and the code number of the last first frame image and the 17th frame The difference from the code number of the image is placed at the 15th position of the newly created data string. When the calculation mode designating unit 37 designates time axis vector quantization, the correlation degree computing unit 36 performs time axis vector quantization on the time axis vector sequence rearranged in this way. Processing is performed, and the time axis code number string of the time axis code book is output.

  At this stage, an image of 640 × 480 pixels for 17 frames is expressed by a data string composed of 19200 data expressed by 11 bits, for example. Therefore, when transferring a moving image for 17 frames, the space direction code book and the time axis code book are sent to the decompression side in advance, and then the time axis code book obtained by the above-described series of processing. What is necessary is just to send the list | wrist expressed with the code number sequence, and the data sequence with which the image of the 1st frame was expressed with the code number sequence of the code book for space directions.

  As a result, an image of 17 frames having an information amount of 41779200 bits can be compressed while maintaining the image quality up to a code number string of a time axis codebook having an information amount of 422400 bits, which is approximately 0.01 times. it can.

  On the decompression side, the original image for 17 frames can be reproduced based on the data transferred from the compression side in this way. The process up to dividing the 307200 array data into 16 spatial direction vectors by the reconstructed image generating unit 48 is the same as in the second embodiment. In the present embodiment, the code number of the spatial direction vector quantized for the first frame sent from the compression side and the code constituting each spatial direction vector reproduced by the reproduction image generation unit 48 Addition with the number difference is performed to reconstruct the code number data string.

  For each of the spatial direction vector of the first frame and the spatial direction vector sequence reconstructed into 16 segments, the reproduced image generation unit 48 codes corresponding to individual code numbers constituting the spatial direction vector. The original image is reproduced by searching for and inserting the vector from the spatial direction codebook stored in the spatial direction codebook storage unit 49. Thereby, the original image for 17 frames can be reproduced.

  In the third embodiment, the number of data constituting one vector is 16, and the space direction codebook and the time axis codebook are each composed of 2048 vectors. This is an example given for the sake of clarity in explaining the present embodiment, and it goes without saying that it should not be limited to the numerical values described here.

  In addition, although it has been described here that images for 17 frames are compressed by vector quantization, the number of frames to be compressed is not limited to 17 frames, and the required number of frames is compressed by vector quantization. Needless to say, you can. Although an embodiment using a moving image is shown here, the present invention is not limited to a moving image and can be similarly applied to an audio signal.

(Fourth embodiment)
FIG. 13 is a data flow diagram illustrating a data flow during compression according to the fourth embodiment. The configuration of the data compression / decompression system according to the fourth embodiment is substantially the same as that of the second embodiment, and is not shown here. However, since the processing contents in the spatial direction code rearrangement calculation unit 39 and the reproduction image generation unit 48 are slightly different in the fourth embodiment, this will be described in detail.

  As shown in FIG. 13, in this embodiment as well as in the third embodiment, the image for 17 frames is compressed, but the contents of the difference calculation performed by the spatial direction code rearrangement calculation unit 39. Is different from the above example. That is, in the third embodiment, the difference between the code number of the first frame image and the code number of each of the remaining frames is calculated, but in the fourth embodiment, the frames adjacent to each other are calculated. Calculate the code number difference.

  That is, as shown in FIG. 13, first, a difference is obtained between the code numbers of the same address in the first frame and the second frame, and then the same address is assigned in the second frame and the third frame. Take the difference between the code numbers. Thereafter, the same processing is performed until a difference is obtained between the 16th frame and the 17th frame. Then, the difference results of the blocks indicated by the same address are rearranged in the order of frame images, so that the difference code numbers of the same address blocks distributed in the time axis direction are collectively expressed as one data string.

  Further, the configuration and operation on the decompression side according to the fourth embodiment are substantially the same as those in the third embodiment, but the contents of the addition performed by the reproduction image generation unit 48 are different. That is, in the fourth embodiment, addition is performed between code numbers constituting the spatial direction vectors of frames adjacent to each other, and a data sequence of code numbers is reconfigured with each addition result.

  That is, first, the code number after the spatial direction vector of the first frame sent from the compression side and the code number difference constituting the first spatial direction vector reproduced by the reproduction image generation unit 48 And the spatial direction vector of the image of the second frame is generated.

  Next, the code number constituting the spatial direction vector of the generated second frame image, and the code number difference constituting the second spatial direction vector generated by the reproduced image generation unit 48 Is added to generate the spatial direction vector of the third frame image. By performing the same operation sequentially for the remaining spatial direction vectors, the code number sequence of the spatial direction vectors for 16 frames excluding the first frame is reconstructed.

  In the fourth embodiment, the number of data constituting one vector is 16, and the space direction codebook and the time axis codebook are each composed of 2048 vectors. This is an example given for the sake of clarity in explaining the present embodiment, and it goes without saying that it should not be limited to the numerical values described here.

  In addition, although it has been described here that images for 17 frames are compressed by vector quantization, the number of frames to be compressed is not limited to 17 frames, and the required number of frames is compressed by vector quantization. Needless to say, you can. Although an embodiment using a moving image is shown here, the present invention is not limited to a moving image and can be similarly applied to an audio signal.

  Here, as a method of obtaining the code number difference between frames, the code number difference between adjacent frames is taken, but it goes without saying that it is not always necessary to obtain the code number difference between adjacent frames.

  As described above, according to the second to fourth embodiments, by introducing the time-axis vector quantization process in addition to the spatial direction vector quantization process, data compression is performed within each frame of the moving image. Data compression can be performed not only between frames but also between frames, and a higher compression rate can be obtained.

  In other words, the compression rate when only spatial direction vector quantization is performed on one frame image is about 1 / 11.6, but by introducing time-axis vector quantization, for example, 16 frames of moving images When data compression is performed on the image data, the compression ratio is 16 times (multiple times the number of frames to be compressed), which is 1 / 1855.6.

(Fifth embodiment)
In the present embodiment, a specific example of a technique for obtaining a higher-quality reproduced image when reproducing a compressed image will be described. As a method for compressing image data in the time axis direction, any of the methods described in the second to fourth embodiments may be used, but here, the case of using the method of the second embodiment is used. For example,

  In the second embodiment described above, the 16 spatial direction vectors generated by re-arranging the 307200 spatial direction code number sequences reproduced by the spatial direction code reproduction unit 44 are individually configured. The code vector corresponding to the code number is referenced from the spatial direction code book in the spatial direction code book storage unit 49, and the original image is reproduced by fitting the data of the corresponding pattern image at each block position.

  However, with this method, a reproduced image of the first frame can be obtained with high image quality, but as the frame advances, quantization errors due to vector quantization in the time axis direction become conspicuous, and the image quality deteriorates. come. Therefore, in this embodiment, in order to prevent such image quality deterioration and maintain the same compression rate, the configuration of the decompressed reproduction image generating unit 48 is as shown in FIG.

  In FIG. 14, the time axis code rearrangement calculation unit 48a rearranges the 307200 spatial direction code number sequences reproduced by the spatial direction code reproduction unit 44 in the same manner as in the second embodiment, and 19200 Sixteen spatial direction vectors composed of data strings are generated. In the present embodiment, a time axis shift unit 48b is provided following the time axis code rearrangement calculation unit 48a.

  In the time axis shift unit 48b, the spatial direction decoding unit 48c refers to the code vector corresponding to each code number of the 16 spatial direction vector sequences from the code book in the spatial direction code book storage unit 49, and has referred to it. When inserting the pattern image into each frame image, insert the pattern image by shifting the elements of the spatial vector in the time axis direction for each desired address of the block defined when vector quantization is performed on the compression side. Shift processing is performed.

  That is, the time axis shift unit 48b performs processing for shifting adjacent elements in the time axis direction for each element (each address) in the spatial direction vector. For example, for a specified address, 16 spatial direction code number sequences are arranged in order from the first frame, but for other addresses, 16 spatial direction code number sequences are other than the first frame. The frames are arranged in order from the first frame.

  By doing in this way, the quantization error caused by the vector quantization in the time axis direction can be dispersed in the time axis direction. Therefore, the quantization error that is conspicuous each time the frame progresses can be made inconspicuous, and the image quality of the reproduced image can be improved as a whole.

(Sixth embodiment)
In the fifth embodiment described above, the quantization error is made inconspicuous by fitting the pattern image by shifting it in the time axis direction on the expansion side. In contrast, the quantization error can be made inconspicuous by performing a shift process in the time axis direction on the compression side. In the sixth embodiment, such a method will be described.

Also in the present embodiment, as a method for compressing image data in the time axis direction, any of the methods described in the second to fourth embodiments may be used, but the method of the second embodiment is used here as well. A description will be given taking the case of using as an example.
In the present embodiment, the spatial direction code rearrangement calculation unit 39 in FIG. 9 performs a process of shifting the code number sequence after spatial direction vector quantization in the time axis direction.

  That is, the spatial direction code rearrangement calculation unit 39, for the code number of the block indicated by the same address in the spatial direction code number sequence generated for each frame image by the spatial direction vector quantization by the correlation degree calculation unit 36, Rearrangement is performed so that the elements of the time axis vector are arranged with reference to different frames for each address.

  That is, in the second embodiment described above, when generating the time-axis vector by rearranging the code numbers of the blocks indicated by the same address in the frame image order in the code number sequence after spatial direction vector quantization, The addresses are rearranged on the basis of the first frame, but in the present embodiment, which frame is rearranged as a reference is changed for each address.

  For example, with respect to a certain time axis vector, the code number of the first frame is placed at the position of address 0 of the newly generated time axis vector data string. The code number of the frame is placed at the position of address 0 in the newly created data string.

  By doing so, the quantization error caused by the vector quantization in the time axis direction can be dispersed in the time axis direction as in the fifth embodiment. Therefore, when a compressed image is reproduced, the quantization error that is conspicuous each time the frame advances can be made inconspicuous, and the image quality can be improved as a whole.

  In the above embodiment, the case where the spatial direction vector quantization and the time axis vector quantization are combined is described as an example, but at least the time axis vector quantization may be performed. At that time, the quantization error caused by the vector quantization in the time axis direction is performed by performing the process of shifting the position between the frames (in the time axis direction) as described above in units of pixels or blocks in each frame image. Can be made inconspicuous.

(Seventh embodiment)
FIG. 15 is a functional block diagram illustrating a configuration example of a data compression system according to the seventh embodiment, and FIG. 16 is a data flow diagram illustrating a data flow during compression according to the seventh embodiment. In FIG. 15, the same reference numerals as those shown in FIG. 9 have the same functions, and thus detailed description thereof will be omitted. In FIG. 16, for convenience of explanation, only one frame of the original image is shown, but actually, a plurality of frames are sequentially input from the image input unit 31.

  The DC component detection / removal unit 51 detects a DC component (minimum luminance value of each pixel) from the block for each block when vector quantization is performed on the original image. In addition, by subtracting the DC component detected for each block from all the pixel values in the block, each pixel value of the input original image is represented by an increment from the minimum luminance value for each block. Generated image.

  The correlation calculation unit 36 is configured only with basic patterns stored in advance in the space direction codebook storage unit 53 when the mode for performing spatial direction vector quantization is specified by the calculation mode specifying unit 52. Using the spatial direction codebook, spatial direction vector quantization is performed on the input image block from which the DC component has been removed. Here, the basic pattern refers to a pattern in which the luminance value of each pixel changes monotonously in eight directions within a code vector block, for example.

  In other words, the basic pattern codebook is, for example, one of the edge portions (upper and lower left and right sides and four corners) of a block configured in units of 4 × 4 pixels. A set of code vectors having a pattern in which luminance values gradually change in an oblique direction. By setting the luminance value of the start point of these code vectors to 0, it is suitable for the image block from which the DC component generated by the DC component detection / removal unit 51 is removed. By giving variations to the increment of the luminance value from the start point to the end point, for example, the spatial direction code book of this basic pattern is constituted by, for example, 321 pattern code vectors.

  Further, the DC component detection / removal unit 51 extracts the minimum luminance value detected for each block. As a result, as shown in FIG. 16, an image of 160 × 120 pixels composed only of the minimum luminance value of each block is generated. Thus, by performing the extraction process of the minimum luminance value by the DC component detection / removal unit 51 and the spatial direction vector quantization process by the correlation degree calculation unit 36, each image of the plurality of frames input as the original image On the other hand, the data string of the spatial direction code number and the data string of the minimum luminance value of each block are generated.

  Next, the spatial direction code rearrangement calculation unit 39 assigns the code number of the block indicated by the same address to the image of each frame represented by the code number sequence of the spatial direction code book by the correlation calculation unit 36. By rearranging in the order of frame images, a time-axis vector in which the code numbers of the same address blocks distributed in the time-axis direction are collected as one data string is generated. This processing may be performed by any method described in the second to fourth embodiments.

  Similarly, the spatial direction DC component rearrangement calculation unit 54 performs the data value of the block indicated by the same address on the image of each frame expressed by only the minimum luminance value of each block by the DC component detection / removal unit 51. Are rearranged in the order of frame images to generate a time-axis vector in which the minimum luminance values of the same address blocks distributed in the time-axis direction are collected as one data string. This processing may be performed by any of the methods described in the second to fourth embodiments.

  In the case of the example shown in FIG. 16, for the time axis vector of the basic pattern, for example, the code numbers of the blocks specified by the head addresses of the first to sixteenth frames are rearranged (127, 287). , 58,..., 283, 276) corresponds to a time axis vector of one basic pattern. For the time axis vector of the minimum luminance value, for example, the minimum luminance values of the blocks specified by the head addresses of the first to sixteenth frames are rearranged (60, 50, 58,..., 76, 77) corresponds to a time axis vector of one minimum luminance value.

  When rearrangement is performed by the spatial direction code rearrangement calculation unit 39 and the spatial direction DC component rearrangement calculation unit 54, a time-axis vector quantization start signal is output and supplied to the calculation mode designating unit 52. In response to this, the calculation mode designating unit 52 switches to a mode in which time-axis vector quantization is performed for either the basic pattern or the minimum luminance value.

  When the calculation mode specifying unit 52 specifies that the time axis vector quantization of the basic pattern should be performed, the correlation degree calculating unit 36 applies each time axis vector generated by the spatial direction code rearrangement calculating unit 39 to each time axis vector. On the other hand, vector quantization is performed using the time-axis codebook stored in the time-axis codebook storage unit 55. Further, when it is specified that the time axis vector quantization of the minimum luminance value is performed by the operation mode specifying unit 52, for each time axis vector generated by the spatial direction DC component rearrangement operation unit 54, Vector quantization is performed using the time-axis codebook stored in the time-axis DC component codebook storage unit 56.

  Through the above processing, a plurality of time axis vectors related to the basic pattern represented by the code number string after space direction vector quantization are replaced with code number strings corresponding to the code vectors in the time axis code book of the basic pattern. At the same time, the plurality of time axis vectors related to the minimum luminance value represented by the data string of the minimum luminance value of each block is replaced with a code number sequence corresponding to each code vector in the time axis code book of the minimum luminance value. These data lists are transferred to the decompression side.

  In the present embodiment, the configuration on the decompression side and the data flow are not shown, but the relationship is basically opposite to that on the compression side. For example, the processing may be performed as follows. . That is, the same expansion process as described in the second to fourth embodiments is performed for each of the basic pattern and the minimum luminance value. When the decompression process is completed for each, the original image can be reproduced by synthesizing the results for each block.

  As described above in detail, according to the seventh embodiment, one feature amount data (minimum luminance value for each block in the above example) is extracted from the original image, and vector quantization of the extracted feature amount data is performed. And the vector quantization of the basic pattern after feature removal, the code used for vector quantization of each image compared to the codebook for the original image itself in which various features are intertwined The book can be simplified, and the number of necessary patterns can be reduced (the 2048 patterns in the second to sixth embodiments are 321 patterns in the present embodiment).

  As a result, when searching for a pattern with a high degree of similarity in the code book by vector quantization, it is easy to narrow down to a more appropriate pattern because of the comparison between simplified patterns. Quantization can be performed. Therefore, when the data compressed in this way is reproduced, a higher quality reproduced image can be obtained.

In the seventh embodiment, the minimum luminance value is described as an example of the feature quantity to be extracted, but the present invention is not limited to this. For example, the maximum luminance value or average value of each pixel in the block may be extracted.
Also in this seventh embodiment, one correlation degree calculation unit 36 is reused for space direction vector quantization and time axis vector quantization, but a plurality of them may be provided for each vector quantization. good.

  Further, the present embodiment can be applied regardless of a monochrome image or a color image. For example, in the case of a color image composed of a Y signal (luminance signal) and U and V signals (color signals), the minimum luminance value, the maximum luminance value, the average value, etc. are extracted as feature amounts from the Y signal as described above. Is possible. Further, not only the luminance signal but also the color signal may be subjected to vector quantization by extracting the minimum value, maximum value, average value, etc. as feature quantities.

  In the above example, only one feature amount is extracted from the original image. However, vector quantization is performed by extracting a plurality of different types of feature amounts (for example, minimum luminance value in a block, direction of luminance change, etc.). May be performed. In this case, the compression rate decreases as the list of code numbers to be output increases, but the quality of the reproduced image can be further improved. If it is desired to prioritize image quality, many feature quantities may be extracted and vector quantized separately. In addition, it may be possible to switch which of the image quality and the compression rate has priority or how much priority.

(Eighth embodiment)
As described above, when compressing a moving image, it is difficult to obtain a high compression ratio by simply performing spatial direction vector quantization on each frame image (still image) constituting the moving image. . In the second to seventh embodiments, the compression ratio is further increased by further performing time-axis vector quantization. On the other hand, in an eighth embodiment described below, means for increasing the compression rate by a method different from the conventional one will be described.

  A configuration example of the data compression system according to this embodiment is shown in FIG. In FIG. 17, an image input unit 61 is a device that continuously captures images. Here, for example, an image of 640 × 440 pixels in which one pixel has an 8-bit gradation is captured at a rate of 30 sheets per second. The image input unit 61 can be realized not only by an image capturing camera or a video camera, but also by a storage medium such as a semiconductor memory device capable of storing images, a hard disk, or a flexible disk.

  A large number of pattern images (code vectors) made up of, for example, 4 × 4 pixel blocks are registered in the code book storage unit 62, and a unique code number is assigned to each block.

  The code book type compression unit 63 performs the spatial direction vector quantization processing described below on each image obtained by the image input unit 61. That is, a block of 4 × 4 pixels is sequentially extracted from the upper left corner of the original image as a starting point in the right direction. After taking out to the right end, the take-out position is shifted down by one block, it is taken out again from the left end, and the blocks for the entire screen are taken out by repeating this.

  For each extracted block, a code vector having the most similar pattern is selected from among a large number of code vectors registered in the code book storage unit 62, and a code number corresponding to the selected pattern vector is output. When an image of 640 × 480 pixels is processed, since 19200 blocks are extracted and processed, 19200 code numbers are output.

  The code string storage unit 64 is for storing a code number string output by processing a frame image at a certain time t by the compression unit 63 using the code book method. Further, the code string comparison unit 65 compares the code number string of the image at the certain time t stored in the code string storage unit 64 with the code number string of the image at the next time t + 1, and indicates the same address. The difference absolute value is calculated for each block code number.

  The output unit 66 determines data to be output to the medium 67 (communication channel or recording medium such as a network) based on the absolute difference value given from the code string comparison unit 65. That is, if the absolute difference value for a certain address is greater than a certain threshold value, the address and code number are output to the medium 67. On the other hand, if the absolute difference value for a certain address is smaller than a threshold value, the address and code number are not output to the medium 67.

  Next, FIG. 18 shows a configuration example when the data compression system shown in FIG. 17 is realized by software. FIG. 18 shows the configuration of a codebook type moving image compression processor including a CPU 68. The image input unit 61, the code book type compression unit 63, and the medium 67 shown in FIG. 18 are the same as those shown in FIG.

  In FIG. 18, a memory 69 provided in association with the CPU 68 includes a code book storage area 69a, a code string storage area 69b, a pointer / counter area 69c, and a threshold value storage area 69d.

  The code book storage area 69a stores a code book used in the code book type compression unit 63. The code string storage area 69b sequentially stores a code number string output for each frame from the code book type compression unit 63, and the image at time t is compressed by the code book type compression unit 63. The resulting code number string is stored.

  The pointer / counter area 69c stores a pointer (address) indicating a position where the code number is read / written to / from the code string storage area 69b. The threshold value storage area 69d stores a threshold value serving as a determination criterion for determining data to be transferred to the medium 67 by the output unit 66.

  FIG. 19 is a flowchart showing the procedure of compression processing executed by the CPU 68 of FIG. In FIG. 19, first, in step S21, the code book data is transferred from the code book storage area 69a to the compression unit 63 using the code book method. Next, in step S22, a predetermined threshold value is set in the threshold value storage area 69b. In step S23, the pointer (address) of the pointer / counter area 69c is initialized, and the pointer value is set to, for example, “0”.

  Next, in step S24, one macroblock (for example, a 4 × 4 pixel block) in the original image is read from the image input unit 61. In step S25, the read macroblock is sent to the compression unit 63 based on the codebook method, where spatial direction vector quantization processing is executed. Thereby, a code number string in the spatial direction is obtained as a compressed code.

  Next, in step S26, the code indicated by the current pointer value using the code number output from the code book type compression unit 63 and the code number obtained by the previous processing and stored in the code string storage area 69b. Calculate the absolute value of the difference from the number. In step S27, it is determined whether the calculated difference absolute value is larger than the threshold value stored in the threshold value storage area 69d.

  Here, if the difference absolute value is larger, in step S28, the code number output from the compression unit 63 by the code book method and the value of the pointer at that time are output to the medium 67, and then the step In S29, the code number is stored in the code string storage area 69b.

  On the other hand, if the calculated difference absolute value is smaller than the threshold value stored in the threshold value storage area 69d, the process proceeds to step S29 without executing the process of step S28, and is output from the codebook compression unit 63. The code number is stored in the code string storage area 69b. In step S30, the pointer in the pointer / counter area 69c is incremented. In step S31, it is determined whether or not the processing has been completed for all macroblocks in one image.

  If the processing has not been completed for all the macroblocks, the process returns to step S24, and the same processing is performed by fetching one next macroblock. The above-described processing is continued until all the macroblocks in the image obtained from the image input unit 61 are processed. When all macroblocks in the image have been processed, all compression processing for one frame of moving image is completed. By repeating the above processing for each frame sequentially, compression of moving images is executed.

  On the other hand, on the decompression side, the basic process of finding a code vector corresponding to the sent code number from the code book and inserting it into each block position is the same as in each of the embodiments described above. It is. However, in this embodiment, since a code number has not been transferred for a block at a certain address, it is necessary to prepare an alternative code vector or pattern image by some method for such a block.

  For this purpose, for example, as a configuration of the data decompression system, a reproduced image storage unit is provided for storing a frame image reproduced by a decoding process at a certain time t. When a reproduced image at time t + 1 is generated by decoding processing, if a code number is not sent from the compression side for a certain block, the image of that block is stored in the previous frame stored in the reproduced image storage unit. Use the image instead. In this way, in order to identify the block to which the code number has been sent and the block to which the code number has not been sent on the decompression side, the address of the block sent together with the code number is used.

  As described above, in the present embodiment, the spatial direction vector quantization process is performed on each of a plurality of frame images, and a spatial direction code number sequence is generated for each frame image. Then, for the code numbers of the same addresses in the images of a plurality of frames, a change amount (difference absolute value) between frames is obtained, and only when the change amount exceeds a certain threshold value, it is used as transfer data. If it is smaller, the data is not transferred.

  In other words, among the blocks making up a multi-frame image, only the data whose data value has changed greatly between frames is transferred, and the data of the previous frame is used without transferring the data whose change is small. To do. As a result, only the data of the part where the motion is intense is transferred, and in addition to the result of the spatial direction vector quantization in each frame, the data can be compressed also in the frame direction (time axis), It is possible to obtain a high compression rate that could not be obtained only by simple spatial direction vector quantization.

(Ninth embodiment)
Each of the embodiments described above can be similarly applied to a monochrome image and a color image. In the ninth embodiment described below, the image quality and compression rate of the reproduced image on the decompression side can be further improved, particularly when data compression by vector quantization is performed on a color image. is there.

  FIG. 20 is a block diagram illustrating a configuration example of the data compression system according to the present embodiment. In FIG. 20, the color original image input from the image input unit 71 is temporarily stored in the image memory 72, read out by the address control or the like from the reading unit 73, and output. The color image read from the image memory 72 is given to the Y, U, V signal separation unit 74.

  The Y, U, V signal separation unit 74 separates the Y signal (luminance signal), the U signal, and the V signal (color signal) from the original image. The blocking unit 75 divides each image data separated by the Y, U, V signal separating unit 74 into blocks of 4 × 4 pixels, for example. At this time, the blocking unit 75 informs the spatial direction codebook storage unit 76 as a signal type designation signal which signal of Y, U, and V is being blocked.

  The correlation calculation unit 77 calculates the above-described similarity independently for each of the Y signal, U signal, and V signal that are separated and extracted in this way. The code determination unit 78 outputs the code number corresponding to the code vector having the highest similarity as the compressed code (Winner code) of the block based on the calculation result of the similarity by the correlation calculation unit 77.

  When performing such spatial direction vector quantization, the correlation degree calculation unit 77 performs vector quantization using a code book stored in advance in the spatial direction code book storage unit 76. Which code book in the book storage unit 76 is used is determined according to the signal type designation signal. That is, the space direction codebook storage unit 76 stores in advance a number of codebooks for Y signals, codebooks for U signals, and codebooks for V signals. The book is sequentially used in accordance with the signal type output from the blocking unit 75 to the correlation degree calculating unit 77.

  At this time, in this embodiment, a codebook larger than the UV signal is assigned to the Y signal. For example, the code book for Y signal is composed of 8192 code vectors (13 bits), and the code book for UV signal is composed of 128 code vectors (7 bits). However, the size of the entire code book including Y, U, and V is set to be the same as or smaller than the overall size when the code books for the respective signals are allocated equally.

  In general, the most important signal that determines the image quality is the Y signal (luminance signal). Therefore, the code book (code length of the code vector) assigned to each signal should be larger for the Y signal than for the UV signal. As a result, the quality of the reproduced image can be improved without changing the overall code length.

  That is, when compression is performed with a compression rate of Y: U: V = 1: 1: 1, the overall code length does not change even if the number of code vectors is assigned to the Y signal more than the UV signal. Therefore, the compression rate for one image does not change. In addition, since the number of code vectors for the Y signal is large, a code vector having a higher similarity can be obtained by vector quantization for the Y signal having a higher importance for the image quality, and the quality of the reproduced image is improved. To do.

  The same effect can be obtained not only when the compression rate is Y: U: V = 1: 1: 1, but also when the compression rate is 4: 1: 1 or 4: 2: 2. For example, when the compression rate is Y: U: V = 4: 1: 1, a 13-bit code is assigned to each block of the Y signal, whereas four blocks are assigned to the UV signal. A 7-bit code is assigned. However, there is no particular problem with respect to color reproducibility even in this way because there is a relatively insensitive property regarding color errors as human visual characteristics.

In the above embodiment, one correlation degree calculation unit 77 is used as a spatial direction vector quantization unit for each of the Y, U, and V signals. However, the spatial direction vector quantization unit is used for each signal. A plurality of signals may be provided, and each signal may be processed in parallel.
In the above embodiment, the code book for Y, U, and V signals is stored in one spatial direction code book storage unit 76, but each code book is divided into different storage units. You may make it memorize.

(Tenth embodiment)
Next, a method of creating a spatial direction code book used when performing spatial direction vector quantization on one image will be described. In this embodiment, the well-known Kohonen learning algorithm is used. FIG. 21 is a diagram showing the procedure of this Kohonen learning algorithm.

  In this method, as shown in FIG. 21, an initial codebook to be given at the time of learning is prepared, and one sample image used for learning is prepared. In learning, a code vector that is most similar to the input image block is obtained by sequentially inputting, for example, 4 × 4 pixel unit image blocks from this sample image and performing spatial direction vector quantization using a code book. Is detected. Based on this result, the contents of the code book are rewritten.

  The rewriting of the code book is performed by determining the gain k and the rewriting range Nc in the mathematical expression representing the Kohonen learning algorithm to appropriate values. By repeating such code vector detection and rewriting processing a predetermined number of times, an optimized code book is obtained. In the present embodiment, various patterns are used as the initial code book when performing such learning.

  FIG. 22 shows, as an initial codebook, a pattern (a) in which luminance values continuously change from a minimum value to a maximum value of possible values, a pattern (b) in which luminance values change randomly, and luminance values A constant pattern (c) is input at 128 (intermediate value), and the codebook is optimized using these. And it is the figure which showed the difference of the PSNR (peak signal to noise ratio) characteristic when an image is compressed and expanded by vector quantization using the codebook obtained thereby. Each of the small squares shown in (a) to (c) represents a code vector. From FIG. 22, it can be seen that a better result is obtained when a pattern whose luminance value continuously changes is input as an initial codebook.

  In FIG. 23, the range in which rewriting is performed by one learning is applied one-dimensionally on the map (virtual two-dimensional plane) as shown in (a), and the rewriting range is decreased as the number of updates increases. The codebook was created with the method of going through and the method of reducing the rewrite range with increasing the number of updates by applying the rewriting range as shown in (b) two-dimensionally, and further compressing / decompressing the image It is the figure which showed the difference of the PSNR characteristic at the time.

  Each of the white circles and black circles shown in FIGS. 23A and 23B indicates a code vector block, and the black circle block indicates a current vector quantization target block. A square range surrounding these white circles and black circles represents a codebook rewrite range. Moreover, the arrow represents the direction which reduces the rewriting range. FIG. 23 shows that the case (a) is superior to the case (b).

  FIG. 24 shows the case where the value of gain k in Kohonen's learning equation is not decreased from the initial value regardless of the increase in the number of codebook updates, and the case where the value is decreased from the initial value as the number of updates is increased. It is a figure which shows the difference in a PSNR characteristic. From FIG. 24, it can be seen that the case where the value of the gain k is decreased from the initial value as the number of code book updates is increased is superior.

  FIG. 25 is a diagram showing the difference in PSNR characteristics when various values are given as the initial value of the gain k and the codebook is created with the number of updates being 160000, for example. From FIG. 25, it can be seen that an excellent result is obtained when the initial value of the gain k is 0.3 to 1.

  From the results shown in FIGS. 22 to 25 described above, the value of the parameter that improves the image quality is identified from various parameters and learning is performed, and the codebook generated thereby is used in the data compression / decompression system. As a result, the quality of the reproduced image can be improved. For example, as an initial codebook given at the time of learning, a pattern in which the luminance value continuously changes from the minimum value to the maximum value that can be taken is used. In addition, the range affected by one learning is applied one-dimensionally, and the range is reduced as the number of updates increases. Furthermore, by setting the initial value of the gain coefficient k given during learning to any value within the range of 0.3 to 1, and decreasing the coefficient value as the number of updates increases, a highly accurate codebook can be obtained. This can be used to improve the quality of a reconstructed image that has been compressed and decompressed.

(Eleventh embodiment)
In the present embodiment, a plurality of code books are created by learning, and a new code book is created by synthesizing them, thereby obtaining a highly accurate code book. FIG. 26 shows a method of combining the characteristics of the three types of code books A, B, and C obtained by optimization into one code book A + B + C.

  In FIG. 26, the size of the code book A + B + C (the number of code vectors) is, for example, 1024 like the code book sizes of A, B, and C alone. That is, the patterns included in the combined code book A + B + C are synthesized by appropriately thinning out the patterns included in the individual code books A, B, and C, respectively.

  Here, the codebook obtained from Kohonen's self-organizing map has similar patterns close to each other on the map, so this characteristic is utilized to make the codebook A, B, C on the map. Every other pattern is taken out in order from the upper left, and this is reconstructed as a new codebook A + B + C. As a result, a code book A + B + C that retains the characteristics of the respective code books A, B, and C is newly generated.

  FIG. 27 shows the PSNR of an image obtained by performing vector quantization on images A ′, B ′, and C ′ using the codebooks A, B, C, and A + B + C shown in FIG. It is a figure which shows a characteristic. It is assumed that the images A ′, B ′, and C ′ are sample images used when creating the individual codebooks A, B, and C, respectively.

  As can be seen from FIG. 27, the single codebooks A, B, and C show generally good results for the sample images A ′, B ′, and C ′ used to create the respective codebooks. Good results have not been obtained for different types of images. For example, when the code book A is used, good results can be obtained for the sample image A ′, but good results cannot be obtained for the other images B ′ and C ′.

  On the other hand, the PSNR characteristics when the synthesized code book A + B + C is used are comparable to the characteristics when the single code books A, B, and C are used for all the images A ′, B ′, and C ′. And has all the characteristics. In this way, by creating a code book by combining several characteristic images, it is possible to create a code book having resistance to various images. For example, it is possible to create a general code book without increasing the size of the code book memory by creating each code book using images of human faces, landscapes, characters, etc. and combining them it can.

  In the above embodiment, a method of combining three types of codebooks of the same size into one is shown. However, the method is not limited to three types, and two or more types of codebooks are combined into one. It can also be synthesized. It is also possible to combine a plurality of codebooks having different sizes into one.

(Twelfth embodiment)
In order to select the most similar code vector from the code book by vector quantization, a huge amount of calculation is required. For this reason, when designing dedicated hardware for vector quantization, an enormous hardware area is usually required to perform high-speed computation. In the embodiment described below, computation can be performed at high speed without increasing the hardware scale.

  FIG. 28 shows a vector in the case where Manhattan distance is used as a function for obtaining a degree of similarity obtained by quantifying whether or not two input vectors are similar, and the sum of elements of vector data is used as an image feature amount. It is a figure for demonstrating the structure of a quantization apparatus, and a mode that a calculation is abbreviate | omitted. When the Manhattan distance is used as the similarity, the similarity is larger as the distance is smaller.

First, the principle that the calculation can be omitted will be described.
The number of dimensions (number of elements) of the vector data is n, and the total number of template vector (code vector) data in the codebook storage unit 81 is k. At this time,
Input vector data: I = (I ₁ , I ₂ ,..., I _n )
Template vector data: T _m = (T _m1 , T _m2 ,... T _mn )
(M = 1, 2,..., K)
If the similarity between the input vector data I and the template vector data T _m is defined by the Manhattan distance M _m expressed by the following equation (1), the template vector data T _m most similar to the input vector data I is The search operation corresponds to an operation for searching for a value of _m that minimizes the distance M _m among all the Manhattan distances M _m (m = 1, 2,..., K).

Here, as shown in the following equation (2), the absolute value D _m of the difference between the sum of the elements of the input vector data I and the sum of the elements of the template vector data T _m , and the above equation (1) Focusing on the fact that the relationship D _m ≦ M _m holds with the Manhattan distance M _m , the following argument holds.

That is, it is assumed that the Manhattan distance M _m between the template vector data T _m and the input vector data I in the code book storage unit 81 is known. At this time, in the codebook storage unit 81, for the above template vector data T _m input vector data and another any template vector data T _o and I, taking the sum of the elements of each vector data, the It is assumed that the absolute value D _o of the difference between the sums is calculated, and as a result, D _o > M _m .

In this case, M _o > M _m always holds due to the relationship of D _m ≦ M _m described above. Thus, without calculating the Manhattan distance M _o of the template vector data T _o and the input vector data I, only to calculate the absolute difference value D _o, the template vector data T _o is compared to the template vector data T _m input It turns out that it is not similar to the vector data I. That is, the possible exclusion of the template vector data T _o from the search.

Looking at more than just, for the template vector data T _o, it is not necessary to calculate the Manhattan distance M _o, if there is need to calculate the difference absolute value D _o, the process involved in the search, as not decreased appear. However, in the calculation of the difference absolute value D _o, (second term on the right side in the formula (2)) sum oT _mi of a feature amount of the template vector data T _m elements, the feature amount storage unit 82 in advance calculated You can remember it.

Therefore, the calculation of the difference absolute value D _o only needs performing calculation taking the difference of the calculation of the feature quantity of the input vector data I, between this and the feature amount feature quantity stored in the storage unit 82. Further, the calculation of the feature amount of the input vector data I can be performed once for one input vector data. Therefore, the calculation of the difference absolute value D _o can be carried out in calculating the amount of very small compared to the calculation of the Manhattan distance M _o.

  Note that using the sum of the elements of the vector data as the feature amount is equivalent to using the average value of the elements. This is because the average value of elements is only the sum of elements divided by the number of elements.

Next, a specific description will be given with reference to FIG.
In FIG. 28, the input vector data I and the plurality of template vector data T ₁ , T ₂ , T ₃ , T ₄ ,... In the code book storage unit 81 are both 6-dimensional vector data as an example. It does not matter in any dimension. The feature amount storage unit 82 includes feature amounts (for example, sum of elements) ΣT ₁ , ΣT _{2 of the} template vector data T ₁ , T ₂ , T ₃ , T ₄ ,... Stored in the code book storage unit 81. , ΣT ₃ , ΣT ₄ ,... Are stored.

In FIG. 28, the template vector data T ₁ , T ₂ , T ₃ , T ₄ ,... In the code book storage unit 81 and the feature amounts ΣT ₁ , ΣT in the feature amount storage unit 82 located immediately on the left side. ₂ , ΣT ₃ , ΣT ₄ ,... That is, the feature amount of the template vector data T ₁ is ΣT ₁ , the feature amount of the template vector data T ₂ is ΣT ₂ , and so on. The data in the code book storage unit 81 and the feature amount storage unit 82 are sequentially read according to the address from the address counter 88.

In addition, the Manhattan distance calculation unit 83 receives the Manhattan distance M _m (m = 1 to k) between the template vector data T _m indicated by d2 transferred from the codebook storage unit 81 and the input vector data I indicated by d1. ) And Manhattan distance data indicated by d3 is output. Further, the minimum Manhattan distance storage unit 84 stores the minimum distance minM among the Manhattan distances M _m calculated so far according to the update control by the calculation omission determination unit 87.

When the calculation with all the template vector data T _m for one input vector data I is completed, the template vector data corresponding to the one stored in the minimum Manhattan distance storage unit 84 is the most input vector data I. Similar data. Then, the address minA of the template vector data that is most similar (the shortest distance) is given from the address counter 88 to the minimum distance address storage unit 89 according to the update control by the calculation omission determination unit 87 and stored therein. .

  Therefore, when the process is completed for one piece of input vector data I, the address stored in the minimum distance address storage unit 89 corresponds to the code number of the block. This code number may be output every time processing of one input vector data I is completed, or may be output collectively when processing of one entire image is completed.

The feature quantity computing unit 85 computes the sum of each element of the input vector data I and outputs the result. During the calculation of the similarity for at least the same input vector data I, the feature values calculated here are stored. Feature quantity difference computing unit 86, a difference between the feature quantity d4 of the input vector data I obtained in the feature amount computing unit 85, a feature amount d5 template vector data T _m read from the feature storage unit 82 Is calculated and output.

The calculation omission determination unit 87 uses the minimum Manhattan distance d6 known so far output from the minimum Manhattan distance storage unit 84 and the feature amount difference d7 calculated by the feature amount difference calculation unit 86. Thus, it is determined whether or not it is necessary to calculate the Manhattan distance between the template vector data _Tm and the input vector data I.

  Then, the determination result is transmitted to the Manhattan distance calculation unit 83 so that unnecessary calculations are omitted, and predetermined control signals are output to the minimum Manhattan distance storage unit 84, the address counter 88, and the minimum distance address storage unit 89. The update control of each content is performed.

  The contents of the address counter 88 are updated when one process is completed, but the contents of the minimum Manhattan distance storage unit 84 and the minimum distance address storage unit 89 are updated only when a predetermined condition is satisfied. If the value of the address counter 208 is incremented before reading the template vector data, the control signal to the Manhattan distance calculation unit 202 is not necessarily required.

Next, a vector quantization procedure of the vector quantization apparatus configured as described above will be described. The manner in which the calculation is omitted will also be described.
1) First, the feature amount of the input vector data I is calculated by the feature amount calculation unit 85, and the result is stored. This feature amount needs to be calculated only once for one input vector data I.

2) Next, the Manhattan distance calculation unit 83 calculates the Manhattan distance M ₁ between the input vector data I and the first template vector data T ₁ in the codebook storage unit 81. The calculation result is shown in the following formula (3). At this stage, since only one Manhattan distance has been calculated, the minimum Manhattan distance minM found so far is M ₁ . Therefore, the Manhattan distance M ₁ is stored in the minimum Manhattan distance storage unit 84.

3) Next, the feature amount ΣT _{2 of the second} template vector data T _{2 in} the feature amount storage unit 82 and the feature amount of the input vector data I stored in the feature amount calculation unit 85 (265 in this example) The feature value difference calculation unit 86 obtains the absolute value D ₂ of the difference between the feature value and the difference value. The calculation result is shown in the following equation (4).

4) Furthermore, the second template vector data T ₂ and the input vector based on the absolute value D ₂ of the feature amount difference obtained in step 3) and the minimum Manhattan distance minM found so far. The calculation omission determination unit 87 determines whether or not the Manhattan distance M ₂ with the data I needs to be calculated. In this case, since D ₂ <minM (= M ₁ ), the calculation of the Manhattan distance M ₂ between the _second template vector data T ₂ and the input vector data I cannot be omitted.

5) Since it is determined in step 4) that the Manhattan distance M ₂ between the _second template vector data T ₂ and the input vector data I needs to be calculated, the Manhattan distance calculation unit 83 performs this calculation. The result of this calculation is M ₂ = 22, which is smaller than the minimum Manhattan distance minM (= M ₁ ) found so far, so the value of minM is updated to M ₂ . Therefore, it is minM = M _2. That is, the minimum Manhattan distance storage unit 84 stores the Manhattan distance M ₂ related to the second template vector data T ₂ .

6) Next, the absolute difference between the feature amount ΣT _{3 of the third} template vector data T _{3 in} the feature amount storage unit 82 and the feature amount of the input vector data I stored in the feature amount calculation unit 85 The value D ₃ is obtained by the feature amount difference calculation unit 86. The calculation result is D ₃ = 46.

7) Furthermore, the third template vector data T ₃ and the input vector based on the absolute value D ₃ of the feature amount difference obtained in step 6) and the minimum Manhattan distance minM found so far. The calculation omission determination unit 87 determines whether or not the Manhattan distance M ₃ with the data I needs to be calculated. In this case, since D ₃ > minM (= M ₂ ), the calculation of the Manhattan distance M ₃ between the _third template vector data T ₃ and the input vector data I can be omitted.

8) Since it has been found that the calculation of the Manhattan distance M ₃ between the _third template vector data T ₃ and the input vector data I can be omitted by the procedure 7), this calculation is omitted. Then, the feature value difference calculation unit 86 calculates the absolute value D ₄ of the difference between the feature value ΣT _{4 of the fourth} template vector data T _{4 and} the feature value of the input vector data I. The calculation result is D ₄ = 157.

9) The fourth template vector data T ₄ and input vector data I based on the absolute value D ₄ of the feature amount difference obtained in step 8) and the minimum Manhattan distance minM found so far. The calculation omission determination unit 87 determines whether or not the Manhattan distance M ₄ needs to be calculated. In this case, since D ₄ > minM (= M ₂ ), the calculation of the Manhattan distance M ₄ between the _fourth template vector data T ₄ and the input vector data I can be omitted.

  10) Thereafter, the same processing is repeated until m = k.

FIG. 29 is a flowchart showing the procedure of the vector quantization process of the present embodiment described above. Hereinafter, this flowchart will be described.
In FIG. 29, first, input vector data I is input in step S41. In step S42, the sum (feature amount) of the elements of the input vector data I is calculated, and the result is stored.

  Next, in step S43, the number (address) of the template vector data to be searched is initialized to m = 1 and minA = 1, and the value of the minimum Manhattan distance minM found so far is initialized to ∞. . Next, in step S44, it is determined whether or not the value of the address counter m exceeds the number k of template vector data. If not, the process proceeds to step S45.

In step S45, the absolute value D _m of the difference between the search and the feature quantity stored in advance for template vector data T _m of a feature quantity of the input vector data I stored result of the operation. Then, in step S46, the calculated difference absolute value D _m is, determines whether or value of the minimum Manhattan distance minM which have been found to date.

Here, the difference absolute value D _m is equal to or larger than the current value of the minimum Manhattan distance MinM, after incrementing the value of the address counter m in step S50, the process returns to step S44. On the other hand, the difference absolute value D _m is if the current minimum Manhattan distance minM smaller, the process proceeds to step S47, calculates an input vector data I, the Manhattan distance M _m of the template vector data T _m looking.

Then, in step S48, the after the calculated Manhattan distance M _m is determined whether a current minimum Manhattan distance minM above. If so, incrementing the value of the address counter m in step S50, processing in step S44 Return to. On the other hand, the calculated Manhattan distance M _m is if the current minimum Manhattan distance minM smaller, the process proceeds to step S49.

At step S49, the value of the minimum Manhattan distance minM renews the Manhattan distance M _m mentioned above calculation, records searched template vector data number (address) to the minimum distance address storage unit 89. Thereafter, after the value of the address counter m is incremented in step S50, the process returns to step S44.

  As shown in the above procedure, when the vector resonating device having the configuration described in FIG. 28 is used, when performing an operation of searching for template vector data most similar to the input vector data I, it is necessary for the search. The calculation of the similarity can be reduced and the calculation can be performed at high speed.

In FIG. 28, the template vector data T ₁ , T ₂ , T ₃ , T ₄ ,... In the code book storage unit 81 are arranged in an arbitrary order. . When the template vector data is arranged in the order of the feature amount, when searching for the template vector data, it is easy to search from the template vector data having the closest feature amount to the input vector data I. Since two vector data having similar feature values tend to have a high degree of similarity, it is possible to omit the calculation of similarity more by performing a search from template vector data having the closest feature value to the input vector data. .

In this embodiment, the dimension of the vector data handled by the vector quantization apparatus is six dimensions, but the dimension of the vector data may be arbitrary.
In addition, the vector quantization apparatus in FIG. 28 can be realized by dedicated hardware, or can be realized by software by a program on a computer. When implemented by hardware, the newly provided feature value calculation unit 85, feature value difference calculation unit 86, and calculation omission determination unit 87 do not require such a large hardware area.

  In the above embodiment, the feature amount of the vector data is not limited to the sum of the elements of the vector data. For example, if it is desired to use the distribution of vector data, the feature amount calculation unit 85 and the calculation omission determination unit 87 may be changed to those for distributed calculation. If the function for determining the similarity is to be changed to a function other than the Manhattan distance, such as a function for determining the Euclidean distance, the Manhattan distance calculation unit 83, the minimum Manhattan distance storage unit 84, and the calculation omission determination unit 87 may be changed. good.

(13th Embodiment)
In the thirteenth embodiment, an example in which image data is used as an input to the vector quantization apparatus shown in the twelfth embodiment and the vector quantization apparatus of the above embodiment is applied to an image compression apparatus is shown in FIG. This will be described with reference to FIG. First, a general example when image compression is performed using a vector quantization apparatus will be described with reference to FIG.

  In FIG. 30, the input original image 91 is configured by a number of elements called pixels. Each pixel has information such as a luminance value or a color difference signal. An input image block 92 is obtained by extracting a block composed of a plurality of pixels from the input image 91. In the example of FIG. 30, 4 × 4 pixels are selected as the size of the input image block 92, but this size may be anything.

  Since the input image block 92 has a plurality of pixels as described above, the luminance values and the like of each pixel can be collected and used as vector data. This is the input vector data I of the vector quantization apparatus described in FIG.

  Due to human visual characteristics, some input image blocks in the input image 91 may look almost the same. A plurality of input image blocks that look the same can be represented by a smaller number of image blocks. The image block code book 93 has a plurality of image blocks (template vector data) representing many input image blocks on the input image 91. The template vector data is obtained by using the luminance of each pixel of the image block in the image block codebook 93 as vector data.

  In the vector quantization apparatus, the entire input image 91 is divided as image blocks, each image block 92 is used as input vector data, and template vector data similar to the input vector data is searched from the code book 93. The image can be compressed by transferring only the number of the corresponding template vector data. In order to reproduce the compressed image and obtain the reproduced image 94, template vector data corresponding to the transferred number may be read from the code book 93 and applied to the image.

  In the present embodiment, if the input vector data indicated by the image block 92 is extracted from the input image 91, the Manhattan distance is used as a function for obtaining the similarity, and the sum of the elements of the vector data is used as the feature amount. The vector quantization procedure is exactly the same as that of the twelfth embodiment described above (the detailed procedure is omitted here). Therefore, when searching for template vector data similar to the input vector data, it is possible to reduce operations related to the search, and thus it is possible to compress an image at high speed.

  FIG. 31 shows how much the calculation of the Manhattan distance can be omitted when the vector quantization operation is performed on the image data by the above method. FIG. 31 shows a template in which the Manhattan distance must be calculated in each template vector data group for various template vector data groups when outdoor landscape photographs, indoor landscape photographs, and portrait photographs are used as input images. It represents the ratio of vector data.

  According to FIG. 31, it can be seen that the template vector data for which the Manhattan distance from the input vector data must be calculated within the template vector data group is within 14% at most. It is shown.

  In the above example, the luminance value of the pixel is vector data. However, not only the luminance value but also any digitized information of each pixel can be vector data. Alternatively, after the input image block 92 is extracted from the input image 91, various transformations such as discrete cosine transformation may be performed, and vector data may be created from the transformed image to be input to the vector quantization apparatus.

  In addition, when the input vector data is image data, the sum of the elements of the vector data is used as the feature amount, or the DC component when the image data is frequency-decomposed is used. This is because the direct current component is simply the sum of the elements of vector data multiplied by a certain coefficient.

(Fourteenth embodiment)
When image data is used as an input to the vector quantization apparatus shown in the twelfth embodiment, it is possible to use image-specific properties as vector data feature quantities. In this embodiment, the characteristic amount specific to image data when the input of the vector quantization apparatus is image data will be described, and the amount of vector quantization processing can be reduced by combining several feature amounts. It is shown below.

  First, with reference to FIG. 32, a problem in the case where the function for calculating the similarity with respect to image data is a Manhattan distance and the sum of vector elements is used as a feature amount will be described. In FIG. 32, an image block 101 corresponding to input vector data and an image block 102 corresponding to template vector data are substantially the same block. On the other hand, the image block 101 corresponding to the input vector data and the image block 103 corresponding to the other template vector data are completely inverted blocks, and are completely different in appearance.

  Actually, when the luminance values of the pixels of the input image block 101 and the template image blocks 102 and 103 are extracted and the Manhattan distance is calculated using each as the vector data, the template image block 103 is compared with the other template image block 102. It becomes clear that the input image block 101 is not similar.

  However, the total sum of the elements of the vector data, which is the feature amount, is equal between the two template image blocks 102 and 103. This means that a lot of useless similarity calculation remains only by using the sum of the elements of the vector data as the feature quantity.

  As a means for solving this problem, there can be considered a means of using the sum of each element after operating so that the brightness of the pixel is inverted for a part of each element of the vector data as the feature amount.

  FIG. 33 shows how the image block changes when the brightness of some of the elements of the vector data is inverted. As a result of operating the initial image block 111 so that the pixels of the blacked out portion are brightly and darkly inverted using the three types of inversion patterns 112 to 114, the after-inversion image blocks 115 to 117 are obtained. As described above, when the light and dark inversion is performed with different inversion patterns 112 to 114 for one image, the sums 118 to 120 of the luminance values of the pixels of the after-inversion image blocks 115 to 117 are different even if the initial image block is the same. Value.

  In FIG. 34, even when the total sum of the elements of the vector data is equal, after operating a part of the elements of the vector data so that the brightness of the pixels is reversed, the sum of the elements of the vector data is obtained and the feature amount is obtained. Then, an example in which a difference appears in the feature amount is shown. In FIG. 34, the sum of the luminance values of the pixels in the initial image blocks 121 and 122 is exactly the same. Therefore, even if the sum of the elements of the vector data, that is, the sum of the luminance values of the pixels in the block is used as the feature amount, the two cannot be distinguished at all.

  For such two initial image blocks 121 and 122, the black portions of the same inversion pattern 123 are brightly inverted, and the sums 126 and 127 of the luminance values of the pixels are calculated for the inverted image blocks 124 and 125. Both values are completely different.

  In the present embodiment, by using the above, it is possible to further reduce the calculation of useless Manhattan distance. That is, as can be seen from the above example, the calculation of similarity that cannot be omitted when a certain feature value is used may be omitted when another feature value is used. In this case, the calculation amount can be further reduced by using two or more different feature amounts.

  Specifically, a method that uses a feature amount in two steps of narrowing down the range in which similarity must be calculated from template vector data with a certain feature amount and further narrowing down the range with another feature amount, For example, a method using one feature amount and using the one having a larger difference in feature amount for the determination of omitting the operation may be used.

  In the case where an input image block is extracted from an image and used as an input to the vector quantization apparatus, the above-mentioned “summation of vector data after operation of a part of vector data elements so that the brightness of pixels is inverted” In addition to the above, there are some feature values specific to an image. In the following, a method of using information on pixels at the four corners of an image block as a feature value and a method of using a change in pixel value on the image block as a feature value will be described.

  First, a method of using information about pixels at the four corners of an image block as a feature amount will be described. When the image block is extracted from the input image, the feature amount can be obtained from the four corners of the image block for an image in which the pixel value smoothly changes in a certain direction in the block. FIG. 35 is a diagram illustrating a method of using the four corners of the image block as feature amounts.

  In FIG. 35A, the image block 131 changes so that the luminance gradually increases from the lower right to the upper left of the block. By extracting the four corners 133 from the image block 131 and comparing the luminance values of these four pixels, the image block 131 can grasp in which direction the luminance changes.

  Further, the image block 132 and its four corners 134 show an example in which the direction in which the pixel value changes is different from the above, and shows a change in luminance that gradually increases from the right side to the left side of the image block. . The arrows shown in 133 and 134 indicate the magnitude relationship of the luminance of the pixels, and the pixel at the tip of the arrow has a higher luminance. When there is no difference in brightness, it is represented by a line segment.

  In this way, the features of the image block can be captured by comparing the brightness between the pixels at the four corners of the image block and determining the size. Since there are various patterns of brightness between the pixels at the four corners, these patterns can be numbered as feature amounts. That is, as shown in FIG. 35 (b), a number is assigned to each pattern of brightness between the pixels at the four corners, and the number is incorporated as a feature amount in the vector quantization apparatus of the present embodiment. Can do.

  In FIG. 35, the size of the image block is 4 × 4 pixels, but this size may be arbitrary. Further, although the luminance value is used as the pixel value, other pixel values (for example, color signal values) may be used.

  Next, a method of using a change in pixel value on an image block as a feature amount will be described. When several image blocks are cut out from the input image, some of the image blocks have similar pixel value changes. FIG. 36 shows two image blocks 141 and 142 having the same pixel value change period and change rate. The pixel values of these image blocks 141 and 142 change only in the left-right direction.

  The change in the pixel value in the left-right direction of the image block 141 is nothing but the suppression of the amplitude of the change by multiplying the change in the pixel value in the left-right direction of the other image block 142 by a predetermined coefficient. Such a state is referred to as having the same pixel value change mode on the image blocks 141 and 142.

Since there are a large number of pixel value change modes on the image block, it is possible to obtain a feature amount by numbering the change modes. That is, a number can be assigned to each pixel value change mode, and the number can be incorporated into the vector quantization apparatus of the present embodiment as a feature amount.
In FIG. 36, the size of the image block is 4 × 4 pixels, but this size may be arbitrary.

  Finally, a pixel value change on the image block, that is, a numbering mode of change on the image block is used as the first feature value, and a pixel value on the image block is used as the second feature value. The configuration of the vector quantization apparatus according to the fourteenth embodiment in the case of using the sum of the values, that is, the sum of the elements of the vector data will be described with reference to FIG. In FIG. 37, the same blocks as those shown in FIG. 28 are denoted by the same reference numerals. Here, the similarity is obtained from the Manhattan distance. In FIG. 37, the vector data is visually expressed as a corresponding image block.

  The input vector data I is classified according to the change mode of the pixel value on the image block which is the first feature amount. That is, the change mode determination unit 151 uses the first data of the input vector data I based on the vector data information stored in the feature template storage unit 152 that collects typical vector data of the pixel value change mode. Data d8 indicating what number of the template corresponds to the change mode of the pixel value on the image block as the feature amount is output.

  The codebook storage unit 81 that stores the template vector data and the feature amount storage unit 82 that stores the sum of the elements of the vector data that is the second feature amount of the template vector data include the type of the first feature amount. Data is arranged for each. In the example of FIG. 37, since the first feature amount of the input vector data I is determined to be “2”, the first feature amount is “2” from the code book storage unit 81 and the feature amount storage unit 82. Only the portion 153 is selected, and the remaining portion is excluded from the search target.

  In the code book storage unit 81 and the feature amount storage unit 82, only for the portion 153 in which the first feature amount is “2” as described above, the template vector data is processed in the same procedure as in the twelfth embodiment. Search for. That is, the second feature amount of the input vector data I is calculated by the feature amount calculation unit 85, and the feature amount difference calculation unit 86 and the calculation are performed using the result d4 and the feature amount d5 in the feature amount storage unit 82. The omission determination unit 87 determines, among template vector data in which the first feature value in the code book storage unit 81 is “2”, in particular, the calculation of the Manhattan distance is not necessary.

  As described above, according to the fourteenth embodiment, more operations can be omitted as compared with the case where the feature amount is used alone as described in the twelfth embodiment. This is because the template vector data to be searched is limited in advance by the first feature quantity, and the search target is further narrowed down by the second feature quantity. As a result, more calculations can be omitted compared to the case where only one feature amount is used, and calculations can be performed at higher speed.

Further, the template vector data is arranged in a two-dimensional array by arranging not only the first feature quantity but also the second feature quantity in order, so that the input vector data I and the second feature quantity are arranged. Since the search can be started from template vector data having similar amounts, template vector data similar to the input vector data I can be searched more efficiently.
In FIG. 37, the size of the image block is described as an example of 4 × 4 pixels, that is, 16-dimensional when converted to vector data. However, this size is arbitrary.

(Fifteenth embodiment)
FIG. 38 is a diagram illustrating a configuration example of a data compression / decompression system according to the fifteenth embodiment. In this embodiment, in a data compression / decompression system using vector quantization, a system for holding a codebook used at the time of data compression / decompression is prepared independently.

  In FIG. 38, a codebook compression system 161 is a device that compresses data input or stored from the outside using a codebook method. The code book server system 162 includes a code book storage device 171, a code book management device 172, a code book registration device 173, and a code book generation device 174, and has a function of transferring a code book in response to a request from the compression / decompression system. Have.

  The code book 165 transferred from the code book server 162 via the network 164 is used as the code book used in the data compression process in the code book type compression system 161. A code book request 163 to the code book server 162 is transmitted via the network 164 to determine what code book is necessary.

  The code book server 162 executes the following operation as a response to the code book request 163. That is, the code book management device 172 searches the code book data stored in the code book storage device 171 for the code book data suitable for the requested one, and searches the code book data compression system 161 via the network 164. Send to. In the code book storage device 171, several code books are registered and stored in advance by the code book registration device 173.

  If there is no code book that matches the requested one in the code book storage device 171, the code book generating device 174 generates a code book that matches the request from the most similar code book. At this time, it may be generated based on the code book stored in the code book storage device 171 or may be generated completely.

The compressed data generated by the codebook compression system 161 using the codebook 165 transferred as described above is transmitted to the codebook decompression system 167 via the network 166.
The code book 170 transferred from the code book server 162 is used as the code book used in the data expansion process in the code book type expansion system 167. A code book request 168 to the code book server 162 is transmitted via the network 169 to determine what code book is required.

  The contents requested by the code book type decompression system 167 may be the request contents generated by the code book type compression system 161 received together with the compressed data and used. Further, the request contents uniquely generated by the code book type decompression system 167 may be used, and different code books may be used for compression and decompression.

  In response to the code book request 168, the code book server 162 executes the following operation. That is, the code book management device 172 searches the code book stored in the code book storage device 171 for the code book data that matches the requested one, and uses the code book storage system 171 to search for the code book data. To 167.

  If there is no code book that matches the requested code book, the code book generating device 174 generates a code book that matches the request from the most similar code book. At this time, it may be generated based on the code book stored in the code book storage device 171 or may be generated completely.

  As described above, in this embodiment, the code book server 162 for holding the code book used for data compression / decompression by vector quantization is prepared independently. Further, the code book server 162, the data compression system 161 (data transmission side), and the data decompression system 167 (data reception side) are represented by analog telephone lines, digital lines, dedicated Internet lines, communication lines, etc., respectively. Connected through a network.

  When the code book distribution request is given to the code book server 162 from each of the data compression system 161 and the data decompression system 167, the code book server 162 transmits the code book instantaneously via the network. .

  As a result, it is not necessary for each data compression / decompression system to hold a code book, and when the code book is updated, it is only necessary to update the data in the code book server 162, so that very simple data transfer is possible. A system can be realized. In addition, the latest code book can always be received from the code book server 162, and the data quality can be kept good.

(Sixteenth embodiment)
In the sixteenth embodiment, after performing vector quantization (spatial direction vector quantization) for each frame, the code numbers obtained thereby are rearranged to perform vector quantization between frames ( An example in which the above-described embodiment for performing (time-axis direction vector quantization) is applied will be described below.

  In the second to seventh embodiments described so far, code numbers after spatial direction vector quantization are extracted one by one from each frame, and rearranged in the order of frame images to generate a new vector (time Axis vector). On the other hand, in the sixteenth embodiment, the code number extracted from one frame is not one, but a plurality of code numbers are extracted from each frame to generate a new time axis vector. is there.

  FIG. 39 is a data flow diagram for explaining how to rearrange the spatial code numbers according to the present embodiment. As shown in FIG. 39, first, spatial direction vector quantization (SDVQ) processing is performed for each macroblock of 4 × 4 pixel units for each frame of the original image, and each macroblock is converted to a spatial direction code vector. Replace with the code number corresponding to. In the example of FIG. 39, the macro block in the right corner of a certain frame is replaced with the code number “30”, and the three surrounding macro blocks are replaced with the code numbers “15”, “10”, and “20”. It has been shown.

  When such spatial direction vector quantization processing is performed for each frame (four frames in the example of FIG. 39), the same address is obtained from each frame image represented by the code number of the spatial direction codebook. Four pieces of code numbers of the indicated blocks are extracted in units of 2 × 2 blocks and rearranged in the order of frame images. As a result, the code numbers for the four blocks with the same address distributed in the time axis direction are grouped as one time axis vector data string, that is, one macro block for time axis direction vector quantization (TDVQ). Express.

  FIG. 40 is a data flow diagram showing the flow of data during compression when the method of rearranging the spatial direction code numbers according to this embodiment is applied to the second embodiment described above, and FIG. 41 and FIG. These are block diagrams which show the structural example of the data compression / decompression system in that case. 40 is the same as the second embodiment shown in FIG. 11 except that the number of frames to be compressed is 4 and the method of code number rearrangement is different. Omitted.

  41, in the data compression system of the present embodiment, a spatial direction code rearrangement method designating unit 181 is further added to the data compression system according to the second embodiment shown in FIG. 42, in the data decompression system of the present embodiment, a spatial direction code reproduction method designating unit 182 is further added to the data decompression system according to the second embodiment shown in FIG.

  The spatial direction code rearrangement method designating unit 181 sets one code number extracted from one frame when rearranging the spatial direction code numbers to generate a new time axis vector, This specifies whether to extract four code numbers from each frame. In other words, it designates whether a series of vector quantization processes is performed according to the flow of FIG. 11 or according to the flow of FIG.

  On the other hand, when the space direction code reproduction method designating unit 182 takes out the code vector corresponding to the time axis code number from the time axis code book, one code for one time axis code number is obtained. This specifies whether to extract a vector or four code vectors. This designation designates a corresponding reproduction method depending on which rearrangement method is designated by the spatial direction code rearrangement method designation unit 181.

  43 and 44 are data flow diagrams showing the flow of data during compression when the method of rearranging the spatial direction code numbers according to this embodiment is applied to the third embodiment and the fourth embodiment described above. It is. 43 and 44, except that the number of frames to be compressed is 5 and the method of rearranging code numbers is different, the third embodiment shown in FIGS. 12 and 13 and Since it is the same as that of the fourth embodiment, the description is omitted here.

  FIG. 45 is a block diagram showing a configuration example of a data compression system when the method of rearranging the spatial direction code numbers according to the present embodiment is applied to the seventh embodiment described above. 45, in the data compression system of the present embodiment, compared to the data compression system according to the seventh embodiment shown in FIG. 15, a spatial direction code rearrangement method designating unit 181 and a spatial direction DC component rearrangement method designating unit. 183 is further added.

  The spatial direction code rearrangement method designating unit 181 is the same as that shown in FIG. Further, the spatial direction DC component rearrangement method designating unit 183 rearranges the data values of the blocks indicated by the same address in the order of the frame images for the images of the respective frames expressed only by the minimum luminance values of the respective blocks. When generating a time axis vector, it is specified whether one DC component is extracted from one frame or four DC components are extracted from each frame.

  As described above, in the sixteenth embodiment, the configuration is such that the spatial direction code numbers and the DC component rearrangement method are designated, and a plurality of code numbers and DC components are extracted from each frame. By making it possible to generate a new time axis vector, time axis vector quantization can be performed in units of blocks grouped in some spatial direction, further improving the quality of the playback image while maintaining a high compression ratio. Can be made.

  In the example shown above, four spatial direction code numbers are extracted from each of the four frames in units of 2 × 2 blocks to form one macroblock of the time axis vector. Nine blocks may be taken out in units of three blocks. In addition to this, it is possible to arbitrarily set a macroblock with a spatial code number and the number of frames to be compressed.

  FIG. 46 shows the relationship between variously set rearrangement methods of spatial direction code numbers and the compression rate at that time. As can be seen, the lower the table shown in FIG. 46, the higher the compression rate.

(Seventeenth embodiment)
Next, a seventeenth embodiment of the present invention will be described. The seventeenth embodiment is a further improvement of the above-described eighth embodiment.
FIG. 47 is a block diagram showing a configuration example of the data compression system according to the present embodiment. The same blocks as those shown in FIG. 17 according to the eighth embodiment are denoted by the same reference numerals.

  In FIG. 47, an image input unit 61 is a device that continuously captures images. Here, for example, an image of 640 × 480 pixels in which one pixel has RGB gradation of 8 bits each is captured at a rate of 30 sheets per second. The image processing unit 191 converts the RGB image signal captured from the image input unit 61 into an image signal composed of a luminance signal (Y signal) and a color signal (U signal, V signal). By performing such processing, the amount of information can be reduced more efficiently by subsequent vector quantization. In the description of the eighth embodiment shown in FIG. 17, such processing is not mentioned, but it may be performed in the same manner.

  A large number of pattern images (code vectors) made up of, for example, 4 × 4 pixel blocks are registered in the code book storage unit 62, and a unique code number is assigned to each block. These code vectors are arranged, for example, in the order of feature amounts described later (for example, address 0 is the code vector with the smallest feature amount, and the ascending order hereinafter).

  The code book type compression unit 63 performs the spatial direction vector quantization processing described in the eighth embodiment on each of the images obtained by the image processing unit 191. That is, for each macroblock included in the original image, the one with the most similar pattern is selected from among many code vectors registered in the codebook storage unit 62, and the corresponding code number is output. When an image of 640 × 480 pixels is processed, since 19200 blocks are extracted and processed, 19200 code numbers are output.

  The code string storage unit 64 is for storing a code number string output by processing a frame image at a certain time t by the compression unit 63 using the code book method. Further, the feature quantity correlation calculation unit 192 includes a code vector read from the code book storage unit 62 based on the code number sequence of the image at the certain time t (previous frame) stored in the code sequence storage unit 64, and Then, the feature amount (for example, average value) with the code vector selected as the most similar pattern as a result of performing spatial direction vector quantization on the image at the next time t + 1 (current frame) is calculated. The correlation between the feature quantities of the code vector of the macroblock indicated by the same address is calculated.

  The output unit 66 determines data to be output to the medium 67 (communication channel or recording medium such as a network) based on the feature value correlation value given from the feature value correlation degree calculation unit 192. That is, when the correlation value of the feature value obtained for an address of a certain macroblock is smaller than a certain threshold value, the address and the code number corresponding to the current frame supplied from the codebook compression unit 63 are output to the medium 67. To do. On the other hand, if the correlation value obtained for an address of a certain macroblock is larger than a certain threshold value, the code number corresponding to that address is not output to the medium 67 (nothing is output). In this case, on the decompression side, as described in the eighth embodiment, for the macroblock, the image reproduced in the previous frame is used instead.

  Here, among the macroblocks between adjacent frames located at the same position in the time axis direction, the macroblock with a small correlation between frames uses the address and code number as transmission data, and the macroblock with a large correlation does not. However, the form of transmission data is not limited to such an example.

  For example, when the correlation is smaller than that of the previous frame, “1” is set at the head, and then a code number is added for transmission. On the other hand, when the correlation is large, only “0” may be set and transmitted. In this way, on the decompression side, by checking whether the leading value of the transmission data sent for each macroblock is “1” or “0”, the code number is sent to the block. It is possible to distinguish between blocks that are not.

  Alternatively, only the code number may be used as transmission data when the correlation is small compared to the previous frame, while only “0” may be used as transmission data when the correlation is large. In this way, on the decompression side, by checking whether the leading value of the transmission data sent for each macroblock is “0” or not, it is sent as the block where the code number is sent. It can be distinguished from no blocks.

  In any of the three patterns of the transmission data format described above, Huffman coding or run length coding may be further applied to transmission data created according to the magnitude of correlation. In this way, the amount of information can be further compressed.

  As described above, in the present embodiment, spatial direction vector quantization processing is performed on each of a plurality of frames of images. Then, the feature values of the selected code vectors are compared between macroblocks of adjacent frames at the same position in the time axis direction, and if the correlation is large, the code number sent in the previous frame is also changed in the current frame. Since it is used as it is, in addition to the result of space direction vector quantization within each frame, it is possible to compress data in the frame direction (time axis), and moving image data can be compared with still image VQ compression. It is possible to compress at a high compression rate.

  In the present embodiment, the threshold value to be compared with the feature amount correlation value in the output unit 66 is constant, but it may be changed between frames based on a certain function or the like. By doing so, for example, the compression rate can be dynamically changed in accordance with the motion of the image by making the threshold variable according to the intensity of motion of the moving image.

  In the above embodiment, the code string of the previous frame is stored in the code string storage unit 64, but the code vector selected in the previous frame may be stored. In this way, the code vector of the previous frame to be compared with the code vector of the current frame by the feature amount correlation degree calculation unit 192 can be directly read from the storage unit.

  In the above embodiment, the feature quantity of the code vector to be compared is obtained by calculation at the time of the compression operation. However, the feature quantity corresponding to each code vector is stored in the code book storage unit 62 in advance. Also good. In this way, the feature amount calculation process during the compression operation can be omitted, and the processing speed of data compression can be increased.

(Eighteenth embodiment)
Next, an eighteenth embodiment of the present invention will be described. The eighteenth embodiment is a further improvement of the seventeenth embodiment described above.
That is, in the seventeenth embodiment, when the correlation between the feature quantities of the code vector is large, the quality of the reproduced image may be deteriorated over time due to the property that the data of the previous frame is used as it is. . For example, when a state in which a correlation value of a code vector feature amount exceeds a threshold value for a certain macro block continues for several frames, deterioration of image quality cannot be denied. It is also difficult to predict such a situation.

  In order to make such image quality deterioration inconspicuous, in this embodiment, “data” is transmitted to the output unit 66 in the data compression system shown in FIG. 47 regardless of the correlation value of the feature amount. The function of “update” is further provided. The feature amount correlation calculation unit 192 determines whether or not to update data or which part of the frame is used as update data. As a data update method in this embodiment, for example, the following three methods can be considered.

  In the first method, for example, the code sequence of all macroblocks obtained by the codebook compression unit 63 for the frame at a rate of once every several frames, the feature value correlation value of the corresponding code vector is larger than the threshold value. Regardless of whether or not, all data is output to the medium 67 as update data. After the update data is output in a certain frame, the update data is used as a new key frame and the compression in the spatial direction and the time axis direction is performed again from there. In this way, even if the state of using the data of the previous frame as it is continues, the state can be surely cut off by the transmission of the update data, and the deterioration of the image quality over time can be suppressed.

  However, in the first method, in the frame in which update data is transmitted, only the vector quantization in the spatial direction is performed by the codebook compression unit 63, and the compression in the time axis direction based on the magnitude of the correlation value is performed. Therefore, the amount of transmission data becomes the maximum (compression rate when ordinary still image compression is performed). Therefore, in the second method of data update, not all data in the frame is used as update data, but data is updated in units of predetermined blocks.

  FIG. 48 is a diagram for explaining a second method of data update. 48A to 48D show frames at certain times t to t + 3, respectively, and a plurality of small blocks shown in the frame are macroblocks of 4 × 4 pixels. In the present embodiment, for example, as indicated by a thick line frame in FIG. 48, the above description is made with a larger block 200 composed of a plurality of macro blocks (in the example of FIG. 48, 2 vertical × 3 horizontal) as a unit. Update data like this.

  That is, in the frame at time t shown in FIG. 48A, in the block 200 at the upper left of the frame, the code string output from the codebook compression unit 63 is the medium 67 regardless of the threshold value processing of the feature amount correlation value. Output to. On the other hand, in the area other than the block 200, as described in the seventeenth embodiment, the feature amount correlation value threshold processing is performed by the feature amount correlation calculation unit 192 and the output unit 66, and the feature amount correlation value is The code is output to the medium 67 only for the macroblock smaller than the threshold value.

  In the next frame at time t + 1 shown in FIG. 48B, the position of the block 200 to be updated is shifted by one block in the horizontal direction. Then, in the block 200 shifted in this way, the code string output from the codebook compression unit 63 is output to the medium 67 regardless of the threshold processing of the feature amount correlation value. On the other hand, in the area other than the block 200, threshold processing of the feature amount correlation value is performed, and the code is output to the medium 67 only for the macroblock whose feature amount correlation value is smaller than the threshold value.

  In the same manner, the process proceeds while shifting the position of the block 200 as shown in FIGS. 48C and 48D. When the position of the block 200 reaches the lower right of the frame, the process shown in FIG. Return to the state and repeat the process.

  In this way, the state in which the data of the previous frame is continuously used as it is by sequentially performing the data update process in units of the block 200 composed of a plurality of macroblocks is the transmission of update data in units of the block 200. The image quality can be prevented from being deteriorated over time. In addition, in the area other than the block 200, the compression in the time axis direction is performed in addition to the vector quantization in the spatial direction, so that a high compression rate can be ensured.

  In the third method of data update, as shown by the thick line frame in FIG. 49, data update is performed for a plurality of macroblocks 201 that are discretely present in one frame, and discrete positions where this data update is performed are performed. Is a method that changes with time (every frame advances). In the second method described above, a difference in image quality may be conspicuous between the block 200 where data is updated and other areas. However, according to the third method, such a difference is inconspicuous. Has merit.

In the second method and the third method of the present embodiment, the blocks 200 and 201 for updating data are provided in every frame, but may be provided every several frames.
Further, although the present embodiment has been described as an improvement of the seventeenth embodiment, it is also possible to apply the data update process described above to the eighth embodiment.

(Other embodiments)
Each functional block and processing procedure shown in the above-described various embodiments may be configured by hardware, or configured by a microcomputer system including a CPU or MPU, ROM, RAM, and the like, and the operation is performed on the ROM and RAM. You may make it implement | achieve according to the stored work program. In addition, what is implemented by supplying a software program for realizing the function to the RAM so as to realize the function of the function block and operating the function block according to the program is also included in the scope of the present invention. include.

  In this case, the software program itself realizes the functions of the above-described embodiments, and the program itself and means for supplying the program to a computer, for example, a recording medium storing such a program are included in the present invention. Configure. As a recording medium for storing such a program, in addition to the ROM and RAM, for example, a flexible disk, hard disk, optical disk, magneto-optical disk, CD-ROM, CD-I, CD-R, CD-RW, DVD, zip, A magnetic tape or a non-volatile memory card can be used.

  Further, by executing the program supplied by the computer, not only the functions of the above-described embodiments are realized, but also the OS (operating system) or other application software in which the program is running on the computer. Needless to say, such a program is included in the embodiment of the present invention even when the functions of the above-described embodiment are realized.

  Further, after the supplied program is stored in the memory provided in the function expansion board of the computer or the function expansion unit connected to the computer, the CPU or the like provided in the function expansion board or function expansion unit based on the instructions of the program Needless to say, the present invention includes the case where the functions of the above-described embodiment are realized by performing part or all of the actual processing.

Examples of embodiments of the present invention are shown below.
(Example 1)
A data string having at least one or more data is blocked into a vector, a code vector similar to a vector extracted from a compression target is searched from a code book prepared in advance, and a code corresponding to the code vector is output. In the data compression device,
The code book is a code book that is optimized by repeating a predetermined number of update processes using a predetermined learning algorithm,
Data compression comprising block shift means for extracting the blocks constituting the vector to be compressed from the object to be compressed by shifting the spatial position in the horizontal direction at least between blocks adjacent in the vertical direction apparatus.
(Example 2)
2. The data compression apparatus according to aspect 1, wherein the blocks constituting the vector are line blocks in which data positions are arranged one-dimensionally.
(Example 3)
A data string having at least one or more data is used as a vector, and a code vector corresponding to the compression code is searched from a code book having at least one or more code vectors and assigned to the corresponding block position. In a data decompression device that reproduces data,
The code book is a code book that is optimized by repeating a predetermined number of update processes using a predetermined learning algorithm,
A data expansion apparatus comprising block shift means for allocating blocks constituting a code vector searched from the code book by shifting a spatial position in a horizontal direction at least between blocks adjacent in the vertical direction.
(Example 4)
4. The data decompression device according to aspect 3, wherein the blocks constituting the vector are line blocks in which data positions are arranged one-dimensionally.
According to the above first to fourth embodiments, blocks constituting a vector used for vector quantization at the time of compression are extracted from the compression target by shifting the spatial position in the horizontal direction at least between blocks adjacent in the vertical direction. Or, the blocks that make up the code vector found in the codebook at the time of decompression are allocated with the spatial position shifted in the horizontal direction at least between the blocks adjacent in the vertical direction. When image data is used, the block boundary (quantization error) can be made inconspicuous by cutting the continuity of each block boundary on the reproduced image. As a result, a high-quality reproduced image can be obtained while maintaining the compression rate.
(Example 5)
A data string having at least one or more data is blocked and converted into a vector. At the time of data compression, a code vector similar to a vector extracted from a compression target is searched from a code book prepared in advance, and corresponding to it. In the data compression / decompression system for reproducing the original data by searching for a code vector corresponding to the code from the code book and assigning the code vector to the corresponding block position when outputting the code and at the time of data decompression,
The code book is a code book that is optimized by repeating a predetermined number of update processes using a predetermined learning algorithm,
First block shift means for extracting the blocks constituting the vector to be compressed from the object to be compressed by shifting the spatial position in the horizontal direction at least between the blocks adjacent in the vertical direction at the time of the data compression; ,
At the time of the data expansion, at least one of the second block shift means for allocating the block of the code vector found in the code book by shifting the spatial position in the horizontal direction at least between blocks adjacent in the vertical direction A data compression / decompression system comprising:
(Example 6)
6. The data compression / decompression system according to example 5, wherein the block constituting the vector is a line block in which data positions are arranged one-dimensionally.
(Example 7)
Example 5 or Example 6 is characterized in that the block shift by the first block shift means and the block shift by the second block shift means have the same shift amount and the opposite directions. The data compression / decompression system described in 1.
(Example 8)
A data string having at least one or more data is blocked into a vector, a code vector similar to a vector extracted from a compression target is searched from a code book prepared in advance, and a code corresponding to the code vector is output. In the data compression method,
The code book is a code book that is optimized by repeating a predetermined number of update processes using a predetermined learning algorithm,
A data compression method characterized in that the blocks constituting the vector to be compressed are extracted from the object to be compressed by shifting the spatial position in the horizontal direction at least between blocks adjacent in the vertical direction.
(Example 9)
9. The data compression method according to aspect 8, wherein the blocks constituting the vector are line blocks in which data positions are arranged one-dimensionally.
(Example 10)
A data string having at least one or more data is used as a vector, and a code vector corresponding to the compression code is searched from a code book having at least one or more code vectors and assigned to the corresponding block position. In a data decompression method that reproduces data,
The code book is a code book that is optimized by repeating a predetermined number of update processes using a predetermined learning algorithm,
A data decompression method characterized in that blocks constituting a code vector searched from the code book are assigned with a spatial position shifted in the horizontal direction at least between blocks adjacent in the vertical direction.
(Example 11)
11. The data decompression method according to example 10, wherein the blocks constituting the vector are line blocks in which data positions are arranged one-dimensionally.
(Example 12)
A data string having at least one or more data is blocked and converted into a vector. At the time of data compression, a code vector similar to a vector extracted from a compression target is searched from a code book prepared in advance, and corresponding to it. In the data compression / decompression method for reproducing the original data by searching for a code vector corresponding to the code from the code book and assigning the code vector to the corresponding block position at the time of data decompression while outputting the code,
The code book is a code book that is optimized by repeating a predetermined number of update processes using a predetermined learning algorithm,
At the time of the data compression, the blocks constituting the vector to be compressed are extracted from the object to be compressed by shifting the spatial position in the horizontal direction at least between the blocks adjacent in the vertical direction.
At the time of data decompression, the block of the code vector found in the code book is at least between the blocks close in the vertical direction so that the shift amount is the same as that at the time of data compression and the directions are opposite to each other. A data compression / decompression method characterized in that a spatial position is shifted and assigned in the horizontal direction.
(Example 13)
13. The data compression / decompression method according to example 12, wherein the block constituting the vector is a line block in which data positions are arranged one-dimensionally.

It is a block diagram which shows the structure of the data compression apparatus by 1st Embodiment. It is a block diagram which shows the structure of the data expansion | extension apparatus by 1st Embodiment. It is a flowchart which shows operation | movement of the data compression apparatus by 1st Embodiment. It is a flowchart which shows operation | movement of the data expansion | extension apparatus by 1st Embodiment. It is a figure for demonstrating the characteristic by 1st Embodiment using the line block as vector data. It is a figure for demonstrating the characteristic by 1st Embodiment which shifted the position which pastes a line block. It is a figure which shows the time required in order to compress / decompress one image of 640x480 pixels. It is a figure for demonstrating the data flow of the whole vector quantization system. It is a block diagram which shows the structure of the data compression apparatus by 2nd-4th embodiment. It is a block diagram which shows the structure of the data expansion | extension apparatus by 2nd-4th embodiment. It is a data flow figure showing a flow of data at the time of compression by a 2nd embodiment. It is a data flow figure showing a flow of data at the time of compression by a 3rd embodiment. It is a data flow figure showing a flow of data at the time of compression by a 4th embodiment. It is a block diagram for demonstrating 5th Embodiment. It is a block diagram which shows the structure of the data compression apparatus by 7th Embodiment. It is a data flow figure showing a flow of data at the time of compression by a 7th embodiment. It is a block diagram which shows the structure of the data compression apparatus by 8th Embodiment. It is a block diagram which shows the structure in the case of implement | achieving the data compression apparatus by 8th Embodiment by software. It is a flowchart which shows operation | movement of the data compression apparatus by 8th Embodiment. It is a block diagram which shows the structure of the data compression apparatus by 9th Embodiment. It is a figure for demonstrating the code book creation method by 10th Embodiment. It is a figure which shows 10th Embodiment and is a figure for demonstrating the difference in the PSNR characteristic at the time of giving an initial codebook variously. It is a figure which shows 10th Embodiment and is a figure for demonstrating the difference in the PSNR characteristic at the time of changing a rewriting pattern variously. It is a figure which shows 10th Embodiment and is a figure for demonstrating the difference of the PSNR characteristic at the time of changing the direction to which the gain was given variously. It is a figure which shows 10th Embodiment and is a figure for demonstrating the difference of the PSNR characteristic when the initial value of a gain is given variously. It is a figure for demonstrating 11th Embodiment. It is a PSNR characteristic figure for demonstrating 11th Embodiment. It is a block diagram which shows the structure of the data compression apparatus by 12th Embodiment. It is a flowchart which shows operation | movement of the data compression apparatus by 12th Embodiment. It is a figure for demonstrating 13th Embodiment. It is a characteristic view for demonstrating 13th Embodiment. It is a figure for demonstrating 14th Embodiment. It is a figure for demonstrating 14th Embodiment. It is a figure for demonstrating 14th Embodiment. It is a figure for demonstrating 14th Embodiment. It is a figure for demonstrating 14th Embodiment. It is a block diagram which shows the structure of the data compression apparatus by 14th Embodiment. It is a block diagram which shows the structure of the data compression / decompression system by 15th Embodiment. It is a data flow figure for demonstrating how to rearrange the spatial direction code number by 16th Embodiment. It is a data flow figure which shows the data flow at the time of compression at the time of applying the rearrangement method of the spatial direction code number by 16th Embodiment to 2nd Embodiment. It is a block diagram which shows the structural example of the data compression apparatus by 16th Embodiment. It is a block diagram which shows the structural example of the data expansion | extension apparatus by 16th Embodiment. It is a data flow figure which shows the flow of the data at the time of compression at the time of applying the rearrangement method of the spatial direction code number by 16th Embodiment to 3rd Embodiment. It is a data flow figure which shows the data flow at the time of compression at the time of applying the rearrangement method of the space direction code number by 16th Embodiment to 4th Embodiment. It is a block diagram which shows the structural example of the data compression apparatus at the time of applying the method of rearrangement of the space direction code number by 16th Embodiment to 7th Embodiment. It is a figure which shows the relationship between the rearrangement method of a space direction code number, and a compression rate. It is a block diagram which shows the example of 1 structure of the data compression system by 17th Embodiment. It is a figure for demonstrating the 2nd method of the data update process by 18th Embodiment. It is a figure for demonstrating the 3rd method of the data update process by 18th Embodiment.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Image input part 2 Image memory 3 Reading part 4 Blocking part 11 Reproduction | regeneration image generation part 12 Block shift part 35 Spatial direction codebook memory | storage part 36 Correlation degree calculating part 37 Operation mode designation | designated part 38 Code determination part 39 Spatial direction code reproduction Array calculation unit 40 Time axis code book storage unit 44 Spatial direction code reproduction unit 45 Time axis code book storage unit 48 Playback image generation unit 48a Time axis shift unit 49 Spatial direction code book storage unit 51 DC component detection / removal unit 52 Calculation Mode Designation Unit 53 Spatial Direction Codebook Storage Unit 54 Spatial Directional DC Component Rearrangement Calculation Unit 55 Time Axis Codebook Storage Unit 56 Time Axis DC Component Codebook Storage Unit 64 Code String Storage Unit 65 Code String Comparison Unit 66 Output unit 74 Y, U, V signal separation unit 75 Blocking unit 76 Spatial direction Codebook storage section 77 Correlation degree calculation section 78 Code determination section 81 Codebook storage section 82 Feature quantity storage section 83 Manhattan distance calculation section 84 Minimum Manhattan distance calculation section 85 Feature quantity calculation section 86 Feature quantity difference calculation section 87 Calculation omission determination Unit 88 address counter 89 minimum distance address storage unit 151 change mode determination unit 152 feature template storage unit 162 codebook server system 171 codebook storage device 172 codebook management device 174 codebook generation device 181 spatial direction code rearrangement method designation unit 182 Spatial direction code reproduction method designating unit 183 Spatial direction DC component rearrangement method designating unit 191 Image processing unit 192 Feature amount correlation degree computing unit 200 Data update block including multiple adjacent macroblocks 201 Multiple discrete macroblocks Data update block containing a click

Claims (63)

A data string having at least one or more data is blocked into a vector, a code vector similar to a vector extracted from a compression target is searched from a code book prepared in advance, and a code corresponding to the code vector is output. In the data compression device,
The code book is a code book that is optimized by repeating a predetermined number of update processes using a predetermined learning algorithm,
For each data at each time to be compressed that changes along the time axis, a plurality of blocks are extracted from the data, code vectors similar to those vectors are searched from the first code book, respectively, First vector quantization means for outputting a corresponding code string at each time;
The code sequence output at each time by the first vector quantization means is rearranged into a plurality of new vectors, and a code vector similar to each of the plurality of new vectors is obtained. A data compression apparatus comprising: second vector quantization means for searching for each of the second codebooks and outputting a code string corresponding to the second codebook.   Code sequence rearranging means for generating a plurality of new vectors by combining the codes of addresses corresponding to the data at each time among the code sequences output at each time by the first vector quantization means. The data compression apparatus according to claim 1, further comprising: A data string having at least one or more data is blocked into a vector, a code vector similar to a vector extracted from a compression target is searched from a code book prepared in advance, and a code corresponding to the code vector is output. In the data compression device,
The code book is a code book that is optimized by repeating a predetermined number of update processes using a predetermined learning algorithm,
For each data at each time to be compressed that changes along the time axis, a plurality of blocks are extracted from the data, code vectors similar to those vectors are searched from the first code book, respectively, First vector quantization means for outputting a corresponding code string at each time;
Of the code sequences output at each time by the first vector quantization means, the code sequence of data at a certain time is used as a reference code sequence, and between the reference code sequence and the code sequence of data at another time Are rearranged to obtain a plurality of new vectors, and for each of the plurality of new vectors, a code vector similar to each of the vectors is added to the second codebook. And a second vector quantization means for searching for each of them and outputting a code string corresponding to them.   Of the code sequences output at each time by the first vector quantization means, the code sequence of data at a certain time is used as a reference code sequence, and between the reference code sequence and the code sequence of data at another time 4. The data compression according to claim 3, further comprising code sequence rearranging means for taking a difference between codes at corresponding addresses and generating a plurality of new vectors by grouping the corresponding address differences. apparatus. A data string having at least one or more data is blocked into a vector, a code vector similar to a vector extracted from a compression target is searched from a code book prepared in advance, and a code corresponding to the code vector is output. In the data compression device,
The code book is a code book that is optimized by repeating a predetermined number of update processes using a predetermined learning algorithm,
For each data at each time to be compressed that changes along the time axis, a plurality of blocks are extracted from the data, code vectors similar to those vectors are searched from the first code book, respectively, First vector quantization means for outputting a corresponding code string at each time;
Among the code strings output at each time by the first vector quantization means, a result obtained by taking a difference between codes at corresponding addresses between data adjacent in the time axis direction is rearranged to obtain a plurality of Second vector quantization means for finding a new vector, searching each of the plurality of new vectors for a code vector similar to each of the vectors from the second codebook, and outputting a code string corresponding thereto And a data compression apparatus.   Of the code strings output at each time by the first vector quantization means, a difference is taken between codes at corresponding addresses between data adjacent in the time axis direction, and the difference between the corresponding addresses is calculated. 6. The data compression apparatus according to claim 5, further comprising code sequence rearranging means for collectively generating a plurality of new vectors. A data string having at least one or more data is used as a vector, and a code vector corresponding to the compression code is searched from a code book having at least one or more code vectors and assigned to the corresponding block position. In a data decompression device that reproduces data,
A second decoding means for finding out and outputting a code vector corresponding to the code string generated by the data compression apparatus according to claim 1 or 2 from the second codebook;
The code strings constituting the plurality of code vectors output by the second decoding means are rearranged into a plurality of new vectors, and the codes constituting the new vectors for the plurality of new vectors, respectively. A data decompressing apparatus comprising: a first decoding unit that searches for and outputs a code vector corresponding to a column from the first code book. A data string having at least one or more data is used as a vector, and a code vector corresponding to the compression code is searched from a code book having at least one or more code vectors and assigned to the corresponding block position. In a data decompression device that reproduces data,
A second decoding means for searching for and outputting a code vector corresponding to the code string generated by the data compression apparatus according to claim 3 or 4 from the second codebook;
Reordering the result of adding the codes of the corresponding addresses between the code string constituting the plurality of code vectors output by the second decoding means and the reference code string to obtain a plurality of new vectors, And a first decoding means for searching for and outputting a code vector corresponding to a code string constituting the new vector from the first code book, respectively, for each of the plurality of new vectors. Data decompression device. A data string having at least one or more data is used as a vector, and a code vector corresponding to the compression code is searched from a code book having at least one or more code vectors and assigned to the corresponding block position. In a data decompression device that reproduces data,
A second decoding means for searching for and outputting a code vector corresponding to the code string generated by the data compression device according to claim 5 or 6 from the second codebook;
For the code strings constituting the plurality of code vectors output by the second decoding means, the respective results obtained by adding the codes in the order of addresses are rearranged into a plurality of new vectors, and the plurality of new vectors A data decompression device comprising: first decoding means for searching each code vector corresponding to a code string constituting the new vector from the first codebook and outputting the code vector. .   The time base shift means is provided, wherein when generating the plurality of new vectors, elements in the new vector are arranged so as to be shifted in the time axis direction between at least adjacent vectors. Item 7. The data compression device according to any one of Items 1 to 6.   8. A time axis shift unit configured to shift at least adjacent elements in the new vector in a time axis direction when generating the plurality of new vectors. The data decompression device according to any one of 9. A data string having at least one or more data is blocked into a vector, a code vector similar to a vector extracted from a compression target is searched from a code book prepared in advance, and a code corresponding to the code vector is output. In the data compression device,
The code book is a code book that is optimized by repeating a predetermined number of update processes using a predetermined learning algorithm,
Of the data sequence at each time to be compressed that changes along the time axis, the data at the addresses corresponding to the data at each time are grouped into a plurality of vectors, and code vectors similar to those vectors are obtained from the codebook. A data compression apparatus comprising vector quantization means for searching for each and outputting code strings corresponding to them.   13. The time axis shift means according to claim 12, further comprising: a time axis shift unit configured to shift elements in the vector in the time axis direction at least between adjacent vectors when extracting the plurality of vectors. Data compression device. A data string having at least one or more data is blocked into a vector, a code vector similar to a vector extracted from a compression target is searched from a code book prepared in advance, and a code corresponding to the code vector is output. In the data compression device,
The code book is a code book that is optimized by repeating a predetermined number of update processes using a predetermined learning algorithm,
Separating means for extracting at least one feature quantity from the compression target and separating it into feature quantity data and basic pattern data obtained by removing the feature quantity from the compression target;
A data compression apparatus characterized in that vector quantization is performed independently for each of the separated feature data and basic pattern data. For the basic pattern data at each time separated from the object to be compressed that changes along the time axis by the separation means, a plurality of blocks are extracted from the data, and code vectors similar to those vectors are extracted as basic patterns. First vector quantization means for finding out each of the first code book for output and outputting a code string corresponding to each code book at each time;
The code sequence output at each time by the first vector quantization means is rearranged into a plurality of new vectors, and a code vector similar to each of the plurality of new vectors is obtained. Second vector quantization means for respectively searching the second code book for the basic pattern and outputting a code string corresponding to the second code book;
The data string that constitutes the feature quantity at each time separated from the compression target that changes along the time axis by the separation means is rearranged into a plurality of new vectors, and each of the plurality of new vectors And a third vector quantization means for searching for code vectors similar to those vectors from the third code book for feature quantities and outputting a code string corresponding to the code vectors. 14. A data compression apparatus according to 14. A data string having at least one or more data is used as a vector, and a code vector corresponding to the compression code is searched from a code book having at least one or more code vectors and assigned to the corresponding block position. In a data decompression device that reproduces data,
First decoding means for basic patterns for reproducing original basic pattern data based on a code string related to the basic patterns output from the data compression device according to claim 14 or 15,
A second decoding unit for feature amount that reproduces original feature amount data based on a code string related to the feature amount output from the data compression device according to claim 14 or 15,
A data decompression apparatus comprising: synthesis means for synthesizing the output results of the first and second decoding means to reproduce the original data. A data string having at least one or more data is blocked into a vector, a code vector similar to a vector extracted from a compression target is searched from a code book prepared in advance, and a code corresponding to the code vector is output. In the data compression device,
The code book is a code book that is optimized by repeating a predetermined number of update processes using a predetermined learning algorithm,
For the data at each time to be compressed that changes along the time axis, a plurality of blocks are extracted from the data, code vectors similar to those vectors are searched from the code book, and the codes corresponding to them Vector quantization means for outputting a sequence at each time;
For the code string output at each time by the vector quantization means, the correlation between the codes is taken between the corresponding addresses of the data adjacent in the time axis direction, and the code is output only for the address where the correlation is smaller than the predetermined value A data compression apparatus comprising: an output control unit configured to do so. A data string having at least one or more data is blocked into a vector, a code vector similar to a vector extracted from a color image to be compressed is searched from a code book prepared in advance, and a code corresponding thereto In the data compression device that outputs
The code book is a code book that is optimized by repeating a predetermined number of update processes using a predetermined learning algorithm,
A codebook for luminance signals and a codebook for color signals are provided, and more code vectors are assigned to the codebook for luminance signals than to the codebook for color signals. Data compression device.   19. The data compression apparatus according to claim 18, wherein a compression rate of the Y signal that is the luminance signal, the U signal that is the color signal, and the V signal is 4: 1: 1 or 4: 2: 2. . A method of creating a codebook used in vector quantization, which is a set of vectors, which is a data sequence having at least one or more data, and using a sample data and an initial codebook Create a codebook that optimizes the codebook by repeating the process and updating the contents of the specified range on the virtual two-dimensional plane of the codebook according to a predetermined update coefficient for each process In the method
A code book creation method using a code book having a pattern in which data values continuously change from a minimum value to a maximum value of possible values as the initial code book. A method of creating a codebook used in vector quantization, which is a set of vectors, which is a data sequence having at least one or more data, and using a sample data and an initial codebook Create a codebook that optimizes the codebook by repeating the process and updating the contents of the specified range on the virtual two-dimensional plane of the codebook according to a predetermined update coefficient for each process In the method
A method for creating a codebook, characterized in that the range to be updated by the one-time processing is applied one-dimensionally on the virtual two-dimensional plane, and the range is decreased as the number of updates increases. A method of creating a codebook used in vector quantization, which is a set of vectors, which is a data sequence having at least one or more data, and using a sample data and an initial codebook Create a codebook that optimizes the codebook by repeating the process and updating the contents of the specified range on the virtual two-dimensional plane of the codebook according to a predetermined update coefficient for each process In the method
A method for creating a code book, wherein an initial value of the update coefficient is set to any value within a range of 0.3 to 1, and the value of the update coefficient is decreased as the number of updates increases. A method of creating a codebook used in vector quantization, which is a set of vectors, which is a data sequence having at least one or more data, and using a sample data and an initial codebook Create a codebook that optimizes the codebook by repeating the process and updating the contents of the specified range on the virtual two-dimensional plane of the codebook according to a predetermined update coefficient for each process In the method
As the initial code book, a code book with a pattern in which data values continuously change from a minimum value to a maximum value of possible values,
Apply the range to be updated in the one-time process one-dimensionally on the virtual two-dimensional plane, and decrease the range as the number of updates increases,
A codebook creation method, wherein an initial value of the update coefficient is set to any value within a range of 0.3 to 1, and the value of the update coefficient is decreased as the number of updates increases. A method of creating a codebook used in vector quantization, which is a set of vectors, which is a data sequence having at least one or more data, and using a sample data and an initial codebook Create a codebook that optimizes the codebook by repeating the process and updating the contents of the specified range on the virtual two-dimensional plane of the codebook according to a predetermined update coefficient for each process In the method
A codebook creation method, wherein a codebook is optimized independently for a plurality of sample data, and a plurality of obtained codebooks are synthesized to create a new codebook.   The code book is created by the method according to any one of claims 20 to 23 when the code book is optimized independently for the plurality of sample data. How to create the code book described in. Feature quantity storage means for storing a feature quantity obtained in advance for each code vector in the code book;
A feature amount calculating means for obtaining a feature amount of a vector extracted from the compression target;
Based on the feature quantity of each code vector stored in the feature quantity storage means and the feature quantity of the compression target vector obtained by the feature quantity calculation means, the code vector and the compression target vector for each of the code vectors. The data compression according to any one of claims 1 to 6, further comprising an operation omitting unit that determines whether or not to omit the operation for obtaining the similarity. apparatus.   27. The data compression apparatus according to claim 26, wherein the feature amount is a sum, an average value, or a direct current component of values of a data string constituting a vector.   The feature amount is a sum, an average value, or a direct current component of values after operating to invert values of some elements of a data string constituting a vector with reference to an intermediate value of possible values. 27. The data compression apparatus according to claim 26.   27. The data compression apparatus according to claim 26, wherein the feature quantity is a direction of change in a block of data string values constituting a vector.   27. The data compression apparatus according to claim 26, wherein the feature amount is an aspect of a change in a block of data string values constituting a vector. The calculation omitting means includes a difference absolute value calculating means for obtaining a difference absolute value between a feature quantity of a certain code vector and a feature quantity of the compression target vector;
When the difference absolute value obtained by the difference absolute value computing means is larger than the minimum value of the values representing the similarity already obtained for other code vectors, the similarity is calculated for the code code. The data compression apparatus according to any one of claims 26 to 30, further comprising an omission determination unit. Feature quantity storage means for storing different types of feature quantities obtained in advance for each code vector in the code book;
Feature quantity computing means for obtaining different types of feature quantities for vectors extracted from the compression target;
Based on the different types of feature quantities related to each code vector stored in the feature quantity storage means and the different types of feature quantities related to the compression target vector obtained by the feature quantity calculation means, 20. An operation omitting unit for determining whether or not to omit an operation for obtaining a similarity to the compression target vector for any one of claims 1 to 6, 10, 12 to 13, and 17 to 19 The data compression apparatus according to item 1. A data compression / decompression system comprising a codebook server that holds at least one codebook, a data compression system, and a data decompression system,
The code book is a code book that is optimized by repeating a predetermined number of update processes using a predetermined learning algorithm,
The data compression / decompression system, wherein the code book server supplies one of the retained code books in accordance with a request from the data compression system and the data decompression system. The code book server includes code book holding means for holding at least one code book;
Code book generating means for generating a code book meeting a request from the data compression system and the data decompression system;
Code book management means for supplying a code book held in the code book holding means or a code book generated by the code book generation means in response to a request from the data compression system and data decompression system. 34. The data compression / decompression system according to claim 33. A data string having at least one or more data is blocked into a vector, a code vector similar to a vector extracted from a compression target is searched from a code book prepared in advance, and a code corresponding to the code vector is output. In the data compression method,
The code book is a code book that is optimized by repeating a predetermined number of update processes using a predetermined learning algorithm,
For each data at each time to be compressed that changes along the time axis, a plurality of blocks are extracted from the data, code vectors similar to those vectors are searched from the first code book, respectively, A first vector quantization step for outputting a corresponding code string at each time;
A code sequence rearrangement step for generating a plurality of new vectors by combining the codes of the addresses corresponding to the data at each time among the code sequences output at each time in the first vector quantization step; ,
For each of a plurality of new vectors generated in the code sequence rearrangement step, code vectors similar to those vectors are respectively searched from the second code book, and a code sequence corresponding to them is output. A data compression method comprising: a vector quantization step. A data string having at least one or more data is blocked into a vector, a code vector similar to a vector extracted from a compression target is searched from a code book prepared in advance, and a code corresponding to the code vector is output. In the data compression method,
The code book is a code book that is optimized by repeating a predetermined number of update processes using a predetermined learning algorithm,
For each data at each time to be compressed that changes along the time axis, a plurality of blocks are extracted from the data, code vectors similar to those vectors are searched from the first code book, respectively, A first vector quantization step for outputting a corresponding code string at each time;
Of the code sequences output at each time in the first vector quantization step, the code sequence of data at a certain time is set as a reference code sequence, and between the reference code sequence and the code sequence of data at another time A code sequence rearrangement step that takes a difference between the codes of the corresponding addresses and generates a plurality of new vectors by combining the differences of the corresponding addresses;
For each of a plurality of new vectors generated in the code sequence rearrangement step, code vectors similar to those vectors are respectively searched from the second code book, and a code sequence corresponding to them is output. A data compression method comprising: a vector quantization step. A data string having at least one or more data is blocked into a vector, a code vector similar to a vector extracted from a compression target is searched from a code book prepared in advance, and a code corresponding to the code vector is output. In the data compression method,
The code book is a code book that is optimized by repeating a predetermined number of update processes using a predetermined learning algorithm,
For each data at each time to be compressed that changes along the time axis, a plurality of blocks are extracted from the data, code vectors similar to those vectors are searched from the first code book, respectively, A first vector quantization step for outputting a corresponding code string at each time;
In the code sequence output at each time in the first vector quantization step, a difference is taken between codes at corresponding addresses between data adjacent in the time axis direction, and the difference between the corresponding addresses is calculated. A code sequence rearrangement step that collectively generates a plurality of new vectors;
For each of a plurality of new vectors generated in the code sequence rearrangement step, code vectors similar to those vectors are respectively searched from the second code book, and a code sequence corresponding to them is output. A data compression method comprising a vector quantization step.   When generating the plurality of new vectors in the code sequence rearrangement step, elements in the new vector are arranged so as to be shifted in the time axis direction at least between adjacent vectors. Item 38. The data compression method according to any one of Items 35 to 37. A data string having at least one or more data is blocked into a vector, a code vector similar to a vector extracted from a compression target is searched from a code book prepared in advance, and a code corresponding to the code vector is output. In the data compression method,
The code book is a code book that is optimized by repeating a predetermined number of update processes using a predetermined learning algorithm,
Of the data sequence at each time to be compressed that changes along the time axis, the data at the addresses corresponding to the data at each time are grouped into a plurality of vectors, and code vectors similar to those vectors are obtained from the codebook. A data compression method comprising a vector quantization step of searching for each and outputting a code string corresponding thereto.   40. The data compression method according to claim 39, wherein when the plurality of vectors are extracted, elements in the vectors are arranged so as to be shifted in the time axis direction at least between adjacent vectors. A data string having at least one or more data is blocked into a vector, a code vector similar to a vector extracted from a compression target is searched from a code book prepared in advance, and a code corresponding to the code vector is output. In the data compression method,
The code book is a code book that is optimized by repeating a predetermined number of update processes using a predetermined learning algorithm,
At least one feature quantity is extracted from the compression target, separated into feature quantity data and basic pattern data excluding the feature quantity from the compression target, and for the separated feature quantity data and basic pattern data A data compression method characterized by performing vector quantization independently of each other. A data string having at least one or more data is blocked into a vector, a code vector similar to a vector extracted from a compression target is searched from a code book prepared in advance, and a code corresponding to the code vector is output. In the data compression method,
The code book is a code book that is optimized by repeating a predetermined number of update processes using a predetermined learning algorithm,
For the data at each time to be compressed that changes along the time axis, a plurality of blocks are extracted from the data, code vectors similar to those vectors are searched from the code book, and the codes corresponding to them A vector quantization step of outputting a sequence at each time;
For the code sequence output at each time in the vector quantization step, the correlation between the codes is taken between the corresponding addresses of the data adjacent in the time axis direction, and the code is output only for the address where the correlation is smaller than the predetermined value And a data compression method. A data string having at least one or more data is blocked into a vector, a code vector similar to a vector extracted from a color image to be compressed is searched from a code book prepared in advance, and a code corresponding thereto In the data compression method for outputting
The code book is a code book that is optimized by repeating a predetermined number of update processes using a predetermined learning algorithm,
A codebook for luminance signals and a codebook for color signals are provided, and more code vectors are assigned to the codebook for luminance signals than to the codebook for color signals. Data compression method.   44. The data compression method according to claim 43, wherein a compression rate of the Y signal as the luminance signal and the U signal and the V signal as the color signals is 4: 1: 1 or 4: 2: 2. .   Comparing the feature quantity of the vector extracted from the compression target with the feature quantity related to each code vector in the code book, and calculating the similarity between the code vector and the compression target vector based on the comparison result The data compression method according to any one of claims 35 to 44, wherein the data compression method is omitted. A difference absolute value calculating step for obtaining a difference absolute value between a feature amount of a certain code vector and a feature amount of the compression target vector;
When the difference absolute value obtained in the difference absolute value calculation step is larger than the minimum value of the values representing the similarity already obtained for other code vectors, the similarity calculation for the code vector is performed. 46. The data compression method according to claim 45, further comprising an omission determination step.   Similar to the compression target vector for each of the code vectors based on the plurality of types of feature amounts and the plurality of types of feature amounts extracted from the compression target vector using a plurality of different types of feature amounts. 47. The data compression method according to claim 45 or 46, wherein it is determined whether or not the calculation for obtaining the degree is omitted.   In a data compression / decompression system including a code book server that holds at least one code book, a data compression system, and a data decompression system, the code book performs an update process using a predetermined learning algorithm. The code book is optimized by repeating a predetermined number of times, and the code book server supplies one of the held code books in accordance with a request from the data compression system and data decompression system. A method for compressing and decompressing data.   A computer-readable recording medium on which a program for causing a computer to function as each means according to claim 1 is recorded.   25. A computer-readable recording medium having recorded thereon a program for causing a computer to execute the processing procedure of the code book creating method according to claim 21.   33. A computer-readable recording medium in which a program for causing a computer to function as each means according to claim 26 or 32 is recorded.   A computer-readable recording medium having recorded thereon a program for causing a computer to realize the function according to claim 33.   A computer-readable recording medium having recorded thereon a program for causing a computer to execute the steps according to any one of claims 35 to 37, 39, and 42.   46. A computer-readable recording medium on which a program for causing a computer to execute the processing procedure of the data compression method according to claim 41 or 45 is recorded.   49. A computer-readable recording medium on which a program for causing a computer to execute the processing procedure of the data compression / decompression method according to claim 48 is recorded. A data string having at least one or more data is blocked into a vector, a code vector similar to a vector extracted from a compression target is searched from a code book prepared in advance, and a code corresponding to the code vector is output. In the data compression device,
The code book is a code book that is optimized by repeating a predetermined number of update processes using a predetermined learning algorithm,
For the data at each time to be compressed that changes along the time axis, a plurality of blocks are extracted from the data, code vectors similar to those vectors are searched from the code book, and the codes corresponding to them Vector quantization means for outputting a sequence at each time;
For the feature quantity of the code vector found at each time by the vector quantization means, the correlation of the feature quantity is performed between corresponding addresses of data adjacent in the time axis direction, and the correlation is smaller than a predetermined value. A data compression apparatus comprising output control means for outputting only a code.   The output control means outputs the code string output from the vector quantization means for all addresses or a part of addresses regardless of whether the correlation is smaller than a predetermined value at a certain time. 57. A data compression apparatus according to claim 17 or 56, wherein:   Regardless of whether the correlation is smaller than a predetermined value, the partial address for outputting the code string is an address including a plurality of adjacent blocks existing in the data to be compressed, or a plurality of discrete addresses existing in the data to be compressed. 58. The data compression apparatus according to claim 57, wherein the address includes a block.   59. A data compression apparatus according to claim 57 or 58, wherein a partial address for outputting a code string is changed over time regardless of whether the correlation is smaller than a predetermined value. A data string having at least one or more data is blocked into a vector, a code vector similar to a vector extracted from a compression target is searched from a code book prepared in advance, and a code corresponding to the code vector is output. In the data compression method,
The code book is a code book that is optimized by repeating a predetermined number of update processes using a predetermined learning algorithm,
For the data at each time to be compressed that changes along the time axis, a plurality of blocks are extracted from the data, code vectors similar to those vectors are searched from the code book, and the codes corresponding to them A vector quantization step of outputting a sequence at each time;
With respect to the feature quantity of the code vector found at each time in the vector quantization step, the correlation of the feature quantity is performed between corresponding addresses of data adjacent in the time axis direction, and the correlation is smaller than a predetermined value. And a data compression method comprising: an output step for outputting only a code.   In the output step, the code string output in the vector quantization step is output for all addresses or a partial address regardless of whether the correlation is smaller than a predetermined value at a certain time. The data compression method according to claim 42 or 60.   Regardless of whether the correlation is smaller than a predetermined value, the partial address for outputting the code string is an address including a plurality of adjacent blocks existing in the data to be compressed, or a plurality of discrete addresses existing in the data to be compressed. 62. The data compression method according to claim 61, wherein the address includes a block.   The data compression method according to claim 61 or 62, wherein a partial address for outputting a code string is changed with the passage of time regardless of whether the correlation is smaller than a predetermined value.

Images