JP2004312379A - Image compression apparatus, image compression method, and image compression program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an image compression device which improves the image quality in a short calculation time. <P>SOLUTION: The image compression device is provided with; a means for dividing an image into blocks; a means for selecting a direction of the maximum luminance variation in each block in a prescribed direction; a means for approximating the two-dimensional luminance change of pixels in the block to a linear luminance change only in the selected direction; a means for making approximate linear data into a vector having pixel values for deviations from the center of the block as the elements and outputting the vector number specifying the vector closest to the shape of the pixel value change out of a set of preliminarily prepared reference vectors; a means for restoring an original vector from the vector number; a means for generating restored data in a direction perpendicular to the change direction from the vector data being the linear pixel value change in the direction of the maximum change to obtain approximate surface data; a means for obtaining error between an original in the block and approximate surface data and calculating an error value of each pixel in the block; and a means for encoding error values distributed in the block. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to an image compression device that compresses an image, an image compression method, and an image compression program.
[0002]
[Prior art]
The image data has a large data amount, and if it is stored without compression, the performance deteriorates in all of the data use speed and the data storage efficiency. Therefore, when storing and using image data, it is necessary to perform highly efficient compression. When compressing an image, the image is divided into small rectangular areas of equal size, and block encoding for compressing within the area is often performed, but with the most efficient compression ratio among the block encodings One of the compression techniques is vector quantization. The technique of vector quantization is described in Patent Document 1 and the like. Here, the basic part of the image compression procedure in the vector quantization will be described.
[0003]
(1) The image is divided into n × n blocks, and k = n × n pixel values in the block are represented by x from the upper left.₀, X₁, ..., x_k-1And the vector x in the k-dimensional space^ρ= (X₀, X₁, ..., x_k-1The following processing is performed for each of the following.
(2) Vector x^ρ= (X₀, X₁, ..., x_k-1) And m prepared reference vectors y^ρi= (Yⁱ ₀, Yⁱ ₁, ..., yⁱ _k-1) And all vectors and distances (differences) D in the table_i: Di = D (x^ρ, Y^{ρ i}) = Σ | x_j-Yⁱ _j|
(3) Further, yi having the smallest Di is obtained from the values, x is represented by an index i = I representing the value, and output as code data by vector quantization.
By performing such a vector encoding process, for example, considering 4 × 4 pixel block data, 8 bits × 16 values can be compressed to about 12 bits.
[0004]
[Patent Document 1]
JP-A-64-60175
[0005]
[Problems to be solved by the invention]
By the way, although it is possible to compress an image with a high compression rate by vector quantization, it is necessary to compare all the vectors of the vector table prepared for reference with the original block vector, so that it takes a very long time. is there. For example, if the number of vectors in the table is 2048, it is necessary to calculate and add the absolute value of the difference X × Y × 2048 times to block the image of vertical X and horizontal Y pixels. Since the amount of calculation is remarkably large, it is difficult to compress an image such as a television rate in real time at the calculation speed of a currently mainstream processing device. In addition, a simple method that can be used for speeding up is to reduce the number of pixels by sub-sampling or to reduce the types of vectors included in the codebook.However, in this case, the pixels become rough or errors may occur. Matching large vectors results in significant degradation of image quality, making it difficult to increase the speed while maintaining image quality. When it is necessary to store an image with a high frame rate with high image quality, there is a problem that vector quantization, which is said to have a high compression ratio, cannot be applied. Furthermore, if the type of table prepared in advance increases, the amount of calculation increases and the data length output as a result of compression increases, so that irregular (not just straight) edges such as characters have a low appearance frequency, and Can not be included in the table prepared for. For this reason, when data representative of a portion such as a character is searched for when compression is performed, the image is far from the original image, and as a result, the character at the time of decompression is deformed and becomes very difficult to read. .
[0006]
The present invention has been made in view of such circumstances, and an image compression apparatus, an image compression method, and an image compression program capable of performing compression processing in a short calculation processing time and improving image quality are provided. The purpose is to provide.
[0007]
[Means for Solving the Problems]
According to a first aspect of the present invention, there is provided a block dividing unit for dividing an entire image into blocks of n × n (n is a natural number of 2 or more) pixels, and m (m is a natural number of 2 or more) pixels in each of the blocks. A maximum change direction detecting means for calculating an amount of change in luminance and selecting a direction in which the amount of change is the largest, and a primary change only in a direction in which the two-dimensional luminance change of the pixel in the block is output by the maximum change direction detecting means. Approximation data generating means approximating the original luminance change, and the approximate one-dimensional data as a vector having pixel values for each deviation from the block center as an element, and calculating the pixel value change from a set of reference vectors prepared in advance. Vector quantization means for searching for a vector closest to the shape and outputting a vector number for identifying the vector, and a vector for restoring a vector which is original approximate data from the vector number Surface data duplicating means for generating restored data in a direction perpendicular to the change direction from vector data which is a one-dimensional pixel value change in the maximum change direction, and as approximate surface data, Error value calculating means for calculating an error value by subtracting the approximate plane data from the original image and calculating an error value of each pixel in the block; and block coding means for coding an error value distributed in the block And characterized in that:
[0008]
The invention according to claim 2 further comprises coding means for entropy coding the maximum change direction output from the maximum change direction detection means, the output of the vector quantization means and the output of the block coding means. Features.
[0009]
According to a third aspect of the present invention, the block encoding means obtains a value for each block that has calculated the quantitative tendency of the undulation of the error and a value for each pixel that represents the distribution of the undulation, thereby obtaining the surface distribution of the error. When the output of the BTC encoder and the output of the BTC encoder are large and the quantitative tendency of the error is large, the output of the BTC encoder is output as it is. Output determination means for outputting a code indicating that there is no BTC code data.
[0010]
The invention according to claim 4 is a block dividing unit that divides the entire image into blocks of n × m (n, m is a natural number of 2 or more) pixels, and the original image in the block is an n × m dimensional vector, A vector quantization means for searching for a vector closest to the original image vector from a set of reference vectors previously expressed only with a rough expression vector and outputting a vector number indicating this vector, and using the original approximate data from the vector number Vector restoring means for restoring a certain vector, error calculating means for obtaining an error value by subtracting the approximate data from the original image in the block to generate an error block, and vector coding for encoding the error block Means, and encoding means for losslessly encoding the output of the vector quantization means and the vector encoding means. And
[0011]
According to a fifth aspect of the present invention, there is provided a block dividing means for dividing an entire image into blocks of nxm (n, m is a natural number of 2 or more) pixels, and 1 / d times and 1 / d times from the image. (D is an integer equal to or larger than 2) a sub-sampling unit that generates a reduced image, a reduced image vector having each pixel of a block obtained by cutting out the reduced image as an element, and a reduced image vector from a set of reference vectors prepared in advance. , A vector quantization means for outputting a vector number for identifying the vector, a vector restoration means for restoring a vector which is original approximate data from the vector number, and a vector restoration means for obtaining a vector number obtained from the vector restoration means. Enlarged data generating means for enlarging the approximate vector to the original size, and obtaining an error value by subtracting and subtracting the approximate vector from the original image in the block, An error calculator for generating a difference block, a vector encoder for encoding the error block, and an encoder for losslessly encoding the outputs of the vector encoder and the quantizer. And
[0012]
According to a sixth aspect of the present invention, there is provided a block dividing step of dividing the entire image into blocks of n × n (n is a natural number of 2 or more) pixels, and a block dividing process in the m (m is a natural number of 2 or more) direction. Calculating the amount of change in luminance and selecting a direction in which the amount of change is the largest, and a primary change only in the direction in which the two-dimensional luminance change of the pixels in the block is obtained in the maximum change direction detection step. An approximation data generation process approximating the original luminance change, and the approximation one-dimensional data as a vector having pixel values for each deviation from the block center as an element, and the pixel value change from a set of reference vectors prepared in advance. A vector quantization process of searching for a vector closest to the shape and outputting a vector number for identifying the vector, and a vector quantization process for restoring a vector that is original approximate data from the vector number Restoring process, surface data restoring process of generating restoration data in a direction perpendicular to the change direction from vector data which is a one-dimensional pixel value change in the maximum change direction and using the surface data as approximate surface data, and an original image in the block. An error value is obtained by subtracting the approximate surface data from the above, and an error value calculating step of calculating an error value of each pixel in the block, and a block encoding step of encoding an error value distributed in the block. It is characterized by having.
[0013]
The invention according to claim 7 further comprises a coding step of entropy coding the maximum change direction obtained in the maximum change direction detection step, the output of the vector quantization step, and the output of the block coding step. I do.
[0014]
According to an eighth aspect of the present invention, in the block encoding step, a value of each block that calculates a quantitative tendency of the undulation of the error and a value of each pixel representing the distribution of the undulation are obtained by calculating the surface distribution of the error. When the BTC encoding process for performing the encoding and the output of the BTC encoding process have a large quantitative tendency of the error, the output of the BTC encoding process is output as it is, and when the amount of undulation is small, An output determining step of outputting a code indicating that there is no BTC code data.
[0015]
According to a ninth aspect of the present invention, there is provided a block dividing step of dividing an entire image into blocks of n × m (n, m is a natural number of 2 or more) pixels, and an original image in the block is an n × m dimensional vector. A vector quantization step of searching for a vector closest to the original image vector from a set of reference vectors previously expressed only with a rough expression vector and outputting a vector number indicating this vector, and using the original approximate data from the vector number A vector restoration process for restoring a certain vector, an error calculation process for obtaining an error value by subtracting the approximate data from the original image in the block, and an error calculation process for generating an error block, and a vector encoding for encoding the error block And a coding step of losslessly coding an output of the vector quantization step and the vector coding step. And
[0016]
According to a tenth aspect of the present invention, there is provided a block dividing step of dividing an entire image into blocks of nxm (n, m is a natural number of 2 or more) pixels, and 1 / d times and 1 / d times from the image. (D is an integer equal to or greater than 2) a sub-sampling process for generating a reduced image, and a reduced image vector from a set of reference vectors prepared in advance as a reduced image vector having each pixel of a block obtained by cutting out the reduced image as an element. , A vector quantization step of outputting a vector number for identifying the vector, a vector restoration step of restoring a vector that is the original approximate data from the vector number, and a vector restoration step. An enlarged data generation process of enlarging the approximate vector to the original size, and an error value is obtained by subtracting and subtracting the approximate vector from the original image in the block. An error calculation step of generating an error block, a vector encoding step of encoding the error block, and an encoding step of losslessly encoding the outputs of the vector encoding step and the quantization step. And
[0017]
According to an eleventh aspect of the present invention, there is provided a block dividing process for dividing an entire image into blocks of n × n (n is a natural number of 2 or more) pixels, and m (m is a natural number of 2 or more) in each of the blocks. A maximum change direction detection process of calculating a change amount of luminance and selecting a direction in which the change amount is the largest, and a two-dimensional luminance change of a pixel in the block in a primary direction only in a direction output by the maximum change direction detection process. Approximate data generation processing that approximates the original luminance change, and the approximate one-dimensional data is a vector having pixel values for each deviation from the block center as an element, and the pixel break change is calculated from a set of reference vectors prepared in advance. A vector quantization process for finding a vector closest to the shape and outputting a vector number for identifying the vector, and a process for restoring a vector that is original approximate data from the vector number. Torsion restoration processing, surface data restoration processing that generates restoration data in a direction perpendicular to the change direction from vector data that is a one-dimensional pixel value change in the maximum change direction, and sets the restored data as approximate plane data, An error value is obtained by subtracting the approximate plane data from the image, an error value calculation process of calculating an error value of each pixel in the block, and a block encoding process of encoding an error value distributed in the block. Is performed by a computer.
[0018]
According to a twelfth aspect of the present invention, the computer further performs a coding process for entropy coding the maximum change direction obtained by the maximum change direction detection process, the output of the vector quantization process, and the output of the block coding process. It is characterized.
[0019]
According to a thirteenth aspect of the present invention, in the block encoding process, a value of each block that calculates a quantitative tendency of the undulation of the error and a value of each pixel representing the distribution of the undulation are obtained by calculating the surface distribution of the error. When the output of the BTC encoding process for performing the encoding and the output of the BTC encoding process has a large quantitative tendency, the output of the BTC encoding process is output as it is. Output determination processing for outputting a code indicating that there is no BTC code data.
[0020]
The invention according to claim 14 is a block dividing process for dividing the entire image into blocks of n × m (n, m is a natural number of 2 or more) pixels, and the original image in the block is an n × m dimensional vector, A vector quantization process of searching for a vector closest to the original image vector from a set of reference vectors previously expressed only with a rough expression vector and outputting a vector number indicating this vector, and using the original approximate data from the vector number A vector restoration process for restoring a certain vector, an error calculation process for obtaining an error value by subtracting the approximate data from the original image in the block to generate an error block, and a vector encoding for encoding the error block Processing and encoding processing for reversibly encoding the output of the vector quantization processing and the vector encoding processing to a computer. It is characterized by performing.
[0021]
According to a fifteenth aspect of the present invention, there is provided a block dividing process for dividing an entire image into blocks of nxm (n, m is a natural number of 2 or more) pixels, and 1 / d times and 1 / d times the image (D is an integer equal to or greater than 2), a sub-sampling process for generating a reduced image, a reduced image vector having each pixel of a block obtained by cutting out the reduced image as an element, and a reduced image vector from a set of reference vectors prepared in advance. , A vector quantization process of outputting a vector number for identifying the vector, a vector restoration process of restoring a vector that is original approximate data from the vector number, and a vector restoration process obtained by the vector restoration process. An enlarged data generation process for enlarging the approximate vector to the original size, and an error value is obtained by subtracting and subtracting the approximate vector from the original image in the block. Causing a computer to perform an error calculation process for generating an error block, a vector encoding process for encoding the error block, and an encoding process for losslessly encoding the outputs of the vector encoding process and the quantization process. It is characterized.
[0022]
BEST MODE FOR CARRYING OUT THE INVENTION
Hereinafter, an image compression apparatus according to an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing the configuration of the embodiment. In this figure, reference numeral 1 denotes a block divider that divides image data to be compressed into blocks of n × n (n is a natural number of 2 or more) pixels. Reference numeral 2 denotes a maximum change direction detector that measures the luminance change in the m direction (m is a natural number of 2 or more) in the block and selects the direction in which the luminance change amount that is the output is the largest. Reference numeral 20 denotes an approximation data generator that approximates a two-dimensional pixel value distribution in a block as a one-dimensional pixel value for each shift position in the maximum change direction output from the maximum change direction detector 2. A vector 3 is a vector having each pixel value of the one-dimensional pixel value data output from the approximate data generator 20 as an element, searching for a vector closest to the approximate pixel value from a set of reference vectors, and outputting its vector number. It is a quantizer.
[0023]
Reference numeral 4 denotes a vector restoring unit that restores a vector that is the original approximate data from the index number output from the vector quantizer 3. Reference numeral 5 denotes a plane data reconstructor that generates approximate plane data by filling a pixel value change, which is a one-dimensional change, in a maximum change direction with the same value in a direction perpendicular to the change direction. Reference numeral 6 denotes an error calculator that subtracts the restored image from the original image in the block to generate a remaining error block. Reference numeral 7 denotes a block encoder that performs encoding in block units different from vector encoding that encodes blocks of error values. The code 8 represents the maximum change direction output from the maximum change direction detector 2, the vector number output from the vector quantizer 3, and the encoded data output from the block encoder 7 by lossless / variable by entropy coding. This is an entropy encoder that performs long compression. Reference numeral 9 denotes a multiplexer that multiplexes the output of the entropy encoder 8 to generate compressed data.
[0024]
Next, only the change in the direction in which the pixel value changes most on the two-dimensional pixel value plane is approximated by vector quantization, and the error between the vector quantization approximate value and the original data is calculated by the second encoder. The operation of encoding will be described.
First, the block generator 1 divides an input image of 4j × 4k (j and k are natural numbers) pixels into j × k blocks of 4 × 4 pixels. FIG. 2 shows an example of a block obtained as a result. In the following description, the block data shown in FIG. 2 will be described as a specific example.
[0025]
Next, the maximum change direction detector 2 detects pixel value changes (referred to as undulating changes) in six types of directions (directions from 0 o'clock to 5 o'clock shown in FIG. 3) every 30 degrees from a vertical line passing through the center of the block. calculate. As shown in FIG. 4, the amount of undulation change in each direction is divided into areas (in this case, four areas) at a fixed distance from the center in the change direction with respect to a straight line for calculating the change direction, and the average pixel value for each of the four areas Is calculated, the absolute value (| L (i) -L (I + 1) |) of the difference between the average pixel values (L (1) to L (4)) of the adjacent areas is calculated, and the sum (| L ( 1) −L (2) | + | L (2) −L (3) | + | L (3) −L (4) |) is obtained as the amount of undulation change in each direction.
[0026]
In the case of 4 × 4 pixels, the area division corresponds to the directions of six types of vector approximation straight lines as shown in FIG. 5, and is divided into four blocks at regular intervals in a direction orthogonal to each vector approximation straight line. This is performed by preparing a method in advance as shown in FIG. In the example shown in FIG. 5, the areas indicated by A, B, C, and D correspond to four areas. If the block is 8 × 8 pixels, the block is divided into eight. When the pixel value shown in FIG. 2 is calculated by the dividing method shown in FIG. 5, the total value of the area B in the area at 1 o'clock is 49 + 78 + 81 + 93 = 301, and the change at 1 o'clock is | 377 / 4-301 / 4 | + | 301 / 4-283 / 4 | + | 283 / 4-149 / 4 | = 57. FIG. 6 shows the amount of undulation change obtained by the same calculation.
[0027]
Subsequently, the maximum change direction detector 2 selects a straight line having the largest total difference between adjacent pixel values among the four pixel values, and determines the direction Dir of the straight line and the average pixel value L of the cut surface by the straight line. (I) is obtained. Calculating with respect to the value of the block shown in the example, Dir = 2 because the direction of 2 o'clock has the largest change amount.
[0028]
Subsequently, the approximate data generator 20 selects an area division method of the block in the maximum change direction from the area division methods prepared in advance for each change direction as shown in FIG. An average value in the divided area is obtained based on the selected division method, and the average value is output as a value representative of the pixel value at each shift position in the change direction. In this example, since the maximum change direction is 2 o'clock, the average of each area in the division method for the 2 o'clock direction shown in FIG. 5 is calculated, and L (1) = 377/3 and L (2) = 301 / 4, L (3) = 283/4 and L (4) = 149/4. In this example, the approximate data generator 20 uses the same calculation method as the maximum change direction detector 2, but the maximum change direction detector 2 calculates the absolute value of the difference for each pixel adjacent in the change direction, and calculates the maximum change direction. The detector 2 may use a different calculation method such as calculating an average for each area.
[0029]
Next, the vector quantizer 3 sets the obtained L as a four-dimensional vector, performs vector quantization, and selects a vector number Idx closest to L (i) from a reference vector table (not shown). Here, the closest is the smallest value obtained by calculating the sum of the absolute value differences between L and the reference vector. Here, a calculation example when the reference vector table has values as shown in FIG. 7 will be described. For example, the absolute value difference sum of the vector number 32 is | 389 / 4-0 | + | 325 / 4-16 | + | 263 / 4-32 | + | 133 / 4-48 | = 211. The sum of the absolute value differences of y is | 389 / 4-96 | + | 325 / 4-80 | + | 263 / 4-64 | + | 133 / 4-32 | = 5.5. Here, assuming that the vector number y is the sum of the minimum absolute value differences, this y is selected as the vector number Idx.
[0030]
Next, the cross-section shape restorer 4 obtains a reference vector numbered with the vector number Idx from the vector table, and obtains its items L '(1)... L' (4). In this example, since the vector number y is selected, L '(1) = 96, L' (2) = 80, L '(3) = 64, and L' (4) = 32 are obtained.
[0031]
Next, the plane data restoration unit 5 writes L ′ (1), L ′ (2), L ′ (3), and L ′ (4) for the pixel at the position corresponding to the previously selected maximum change direction, respectively, Obtain a vector restoration image. As a result, a calculated intermediate value shown in FIG. 8A is obtained. In this example, as a method of restoring two-dimensional data from one-dimensional data, a NearestNeighbor method that determines a corresponding position and selects a value is used. However, a Bi-Linear method that performs primary interpolation according to a position shift is used. May be used.
[0032]
Next, the error calculator 6 obtains error block data by subtracting the previously obtained vector restored image from the original image block for each pixel at a corresponding position. That is, the intermediate value shown in FIG. 8A is subtracted from the pixel value shown in FIG. 2, and as a result, the value of the calculated intermediate value shown in FIG. 8B is obtained.
[0033]
Next, the block encoder 7 calculates, for each pixel of the error block data, quantized data that is 1 if 0 or more and 0 if less than 0, the number of positive data is NumP, and the average of positive data is AvrP. If the number of negative data is NumN and the average of the negative data is -AvrN, encoding is performed by BlockTruncationCoding centering on 0 by obtaining the amplitude index as (NumP × AvrP + NumN × AvrN) / (NumP + NumN). As a result, a quantization pattern value in which 16 pieces of quantization data (1 bit per pixel) indicating whether to add or subtract an amplitude index for each pixel, a BTC amplitude index (8 bits) D, and compression to 24 bits. Generate data. In this example, the quantization pattern value shown in FIG. 9 is obtained, and the amplitude index 7.5 (= 120/16) can be obtained.
[0034]
Note that the amplitude index obtained by the above calculation is equal to or less than a threshold (3: high sensitivity to 12: low sensitivity, which is set to about 6; usually, about 6) which determines the sensitivity for the detailed expression set in advance. The BTC code of the amplitude index is replaced with a code indicating “no BTC code”. As a result, it is possible to prevent the data from including subtle changes and noises that cannot be distinguished or conscious by humans, and to restore the image having the same appearance while reducing the data amount. In this example, since the amplitude index is 7 and exceeds 6, the previously obtained BTC encoding value is used as the encoding result.
[0035]
Next, the entropy encoder 8 entropy-compresses the four types of data (the maximum change direction Dir, the BTC amplitude index D, the quantization pattern value q, and the vector number Idx) obtained by the above-described block coding. Then, the multiplexer 9 multiplexes the obtained entropy-encoded data and outputs compressed data.
[0036]
Note that the block encoder 7 quantizes the block data using discrete cosine transform, BTC (Block Truncating Coding) method, wavelet transform, or the like, and then performs quantization to generate compressed data. May be. In particular, efficient compression can be performed by using a two-stage (1 bit) BTC transform for the block encoder 7 in the quantization stage. At this time, as shown in FIG. 10, the block encoder 7 obtains “a value for each block in which the quantitative tendency of the undulation of the error is calculated” and “a value for each pixel representing the distribution of the undulation”. A BTC encoder 71 that encodes the surface distribution of the error, and outputs the BTC encoded data as it is when the error tendency (= undulation amount) of the error received from the output of the BTC encoding is large. Is small, it is sufficient to provide an output determiner 72 for adding a code of "no BTC code data" assuming that even if the error portion is decoded, that portion is overlooked.
[0037]
Next, an operation of decoding an image compressed by the above-described operation will be described.
First, the maximum change direction Dir = 2, the BTC amplitude index D = 7, the vector number Idx = y, and the quantization pattern value q = 01011111111010100 (binary number) are restored from the data compressed by the entropy coding.
[0038]
Next, vector data is searched from the vector number, and as shown in FIG. 11, an approximate block is obtained by coloring pixel values sequentially in the maximum change direction. In this example, since the vector number is y, L (1 to 4) = 96, 80, 64, 32 is decoded, and the value to be painted from the value in the maximum change direction 2 is painted in the pattern at 2 o'clock. Since it is decided to divide, FIG. 8A can be restored. In this example, as a method of restoring two-dimensional data from one-dimensional data, a NearestNeighbor method that determines a corresponding position and selects a value is used. However, a Bi-Linear method that performs primary interpolation according to a position shift is used. May be used. However, the same restoring method as that of the plane data restorer 5 is used as a method of restoring plane data at the time of decoding.
[0039]
Next, decoding for the BTC encoding method is performed to obtain an error correction value. In this example, the correction amount block shown in FIG. 12 can be restored from the quantization pattern value q = 0101 1111 1110 0100 (binary number) and the amplitude index = 7. Then, with respect to the block data shown in FIG. 8A and the block data shown in FIG. 12, data at corresponding locations in the block are added to each other, and as a result, a decoded block is obtained. In this example, the block data shown in FIG. 13 is obtained.
[0040]
Next, image data is reconstructed by laying out the decoded results for each block to obtain a restored image.
[0041]
As a result of the above decoding, when the obtained decoding example (FIG. 14) is checked, the inclination of the entire pixel value (thick arrow in FIG. 14) and the sharp The edge change (thick line to double line in FIG. 14) is restored. In an actual image, there is almost no local variation such as 4 × 4 pixels other than the characteristic portion, and the image can be restored more faithfully than the image shown in this example. Since it is important to correctly understand the image when the decoded image is seen, the present method is a lossy compression method that can restore an image that can be correctly read when viewed by a human. .
[0042]
Next, another embodiment will be described with reference to FIG. FIG. 15 is a block diagram showing a configuration of another embodiment. The image compression apparatus according to this embodiment includes a block converter 11 that divides an entire image into blocks of n × m (n and m are natural numbers) pixels, an original image in the block as an n × m dimensional vector, and a reference vector (not shown). A vector closest to the original image vector is searched for from the set of tables, and a vector quantizer 12 that outputs a vector number indicating the vector, and a vector that is original approximate data is restored from the vector number output from the vector quantizer 12 A vector restoring unit 13, an error calculator 14 for subtracting a restored image obtained from the vector restoring unit 13 from the original image in the block to generate a remaining error block, and a block of an error value obtained from the error calculator 14. A block encoder 15 for performing block-based coding different from vector coding for coding It comprises an entropy encoder 16 for performing lossless / variable-length compression of output data from the child encoder 12 and the block encoder 15 by entropy encoding, and a multiplexer 17 for multiplexing the entropy encoded data. In this image compression apparatus, only a vector of a rough expression is recorded in a reference vector table referred to by the vector quantizer 12, and the number is smaller than the vector quantizer 12 required to compress a target image quality. It is characterized in that a (less than 1/4) vector table is prepared.
[0043]
Next, an image compression apparatus according to another embodiment will be described with reference to FIG. FIG. 15 is a block diagram showing a configuration of another embodiment. The difference between the image compression device in this embodiment and the image compression device shown in FIG. 1 is that the image compression device shown in FIG. 1 converts data as a line instead of a surface into a vector in order to reduce the number of dimensions (number of parameters) of the vector. Although the quantization is performed, the image compression apparatus shown in FIG. 16 performs sub-sampling on the surface data to reduce the number of dimensions of the vector using data in which the number of pixels in the vertical and horizontal directions is reduced. The image compression apparatus according to this embodiment includes a block generator 21 that divides the entire image into blocks of nxm (n and m are natural numbers) pixels, and a thinning process or a process that reduces the image data to a 1 / d size. A sub-sampler 22 that performs reduction and generates data 1 / d times vertically and 1 / d times horizontally (d is an integer of 2 or more) and a vector having each pixel of a block obtained by cutting out the reduced image as an element , A vector closest to the reduced image vector from the set of reference vectors, and a vector quantizer 23 that outputs a vector number indicating the vector, and a vector that is original approximate data based on the vector number output from the vector quantizer 23. And an expanded data generator 25 for expanding the approximate vector obtained from the vector recoverer 24 to the original size. An error calculator 26 for subtracting the restored image from the enlarged block generator from the original image in the block to generate a remaining error block, and a block different from the vector coding for coding the block of the error value from the error calculator 26 A block encoder 27 that performs unit encoding, an entropy encoder 28 that performs lossless and variable-length compression of output data of the vector quantizer 23 and the block encoder 27 by entropy encoding, and an entropy encoded data And a multiplexer 17 for multiplexing.
[0044]
Next, comparison results between the image compression method according to the present invention and the image compression method according to the related art will be described.
<Calculation amount>
Compared to compression using only vector quantization, the present invention prepares only a small vector table omitting the vector table taking into account changes in small parts and refers to it, so that the processing time that takes the most time The amount is reduced to a fraction, and the feature of low compression speed, which has been a problem in vector quantization, can be solved. For example, consider a vector coding having a codebook of 2048 vectors with a block size of 4 × 4 pixels, the target is from 16 to 4 parameters, and the change of the codebook itself in the x and y directions is also in the x direction. , The number of types is reduced to at least a half or less, so that the calculation amount can be reduced to 1/8 or less.
[0045]
<Image quality>
In the present invention, the distribution of pixel value fluctuations in a direction orthogonal to a large change that cannot be reproduced by single-direction vector quantization, and a portion including a small spatial change, are obtained from block data approximated by rough vector quantization. , And compresses the error by an encoding method that requires less calculation than another vector quantization. At the time of restoration of the encoded data, since an image equivalent to an image in a code book restored with data of two-dimensional vector encoding can be decoded in each block, the same image is collected and decoded. It will have image quality.
As shown in FIGS. 17A and 17B, the density change in the direction of the arrow having the maximum inclination is encoded / decoded, and the decoding is performed in such a manner that the same color is applied in the direction perpendicular to the arrow, and the BTC encoding is performed. By not adding data, the same image quality as the original image can be obtained.
Also, as shown in FIG. 17C, the density change is encoded and decoded in the direction of the arrow, which is the maximum slope, and the decoding is performed by painting the same color in the direction orthogonal to the arrow. Decode a correction value distribution having a correction value that makes the part shown in (1) darker by a certain amount (difference between white part and dark part) from the BTC code (no error because it is binary information) and sum the above two values Can obtain the same image quality as the original image.
[0046]
In this manner, solid color or gradation (equivalent to FIG. 17A) or regular color unevenness as shown in FIG. 17B, which is easily noticeable locally by humans, and their boundary (FIG. In c)), even if this method is used, an image equivalent to two-dimensional vector quantization can be restored. Further, the vector quantization of the two-dimensional pixels itself includes data having the above-described characteristics in many of the vector data prepared in advance in the vector table as a local form of an image that is understood by humans in a natural image. By obtaining numbers that refer to these, the data is compressed as data expressing only those parts that are easily noticed by humans. Compared with the vector quantization method for such a natural image, even if this method is used, an image for each block is restored at the same level for the same feature. In this way, just as the vector quantization method can achieve the same image quality as the original image in the range that humans can be aware of when restoring, this method also reduces the original image to the extent that humans can be aware of features that mostly occupy natural images. The same image quality as that of the image can be obtained.
[0047]
<Comparison with BTC coding>
The rough structure in the block is compressed by vector quantization, and the block to which the correction is performed and the BTC code is added is limited to a part having a complicated structure in the block, and a part having a complicated structure such as a natural image is In the case of an image having only a small number of bits, the code amount is reduced as compared with the BTC coding. For example, only the vector quantization part is 11 bits + 4 bits in the direction (because the method of painting does not change even if it is finer than 16 steps which are more than 12 pixels on the outer periphery which are 4 × 4 blocks) + the presence or absence of BTC 1 bit, the code amount of BTC Is 32 bits (center color, undulation amount, undulation distribution 16 bits). When BTC is used on the entire surface, each block is 32 bits. In this method, a portion with BTC coding correction can be expressed by 16 + 32 bits, and a portion without correction can be expressed by 16 bits. Thus, when the complexity is less than 50% and the encoding is performed by the present method, the compression ratio is better than BTC. Also, in a normal natural image, a complicated portion is about 10%, and even if it is large, most of it is within about 40%. For this reason, as a coding apparatus for general use of natural pictures (without parameter adjustment), the compression ratio is better than that of BTC. In addition, this method has the same image quality as vector coding, which is said to have the same image quality and better compression rate than BTC. It becomes.
[0048]
<Comparison of image quality of images with edges and complex patterns such as characters with conventional vector quantization>
With regard to edges when characters that cannot be accurately represented by vector quantization are captured as an image, the BTC method is used to block-code error value blocks, and extreme undulations specific to edges are applied to each pixel. It can be expressed by an expression that switches between concave and convex. At this time, the function necessary to approximate the error as a model that reproduces the details of the original image is represented by a minimum amount of information such as two-stage quantization. On the other hand, in order to reproduce characters of all thicknesses in normal vector quantization, it is necessary to prepare a pattern of 2 to the 16th power in which the binary values of the shades are exchanged for all pixels. It is not practical considering the amount of calculation on the order of the 16th power. As described above, the image quality of an image including a character can be improved by using the present method as compared with the vector quantization that cannot be expressed using the vectors corresponding to all the character patterns.
[0049]
A program for realizing the function of each processing unit in FIGS. 1, 10, 15, and 16 is recorded on a computer-readable recording medium, and the program recorded on this recording medium is read by a computer system and executed. By doing so, the image compression processing may be performed. Here, the “computer system” includes an OS and hardware such as peripheral devices. The “computer system” also includes a WWW system provided with a homepage providing environment (or display environment). The “computer-readable recording medium” refers to a portable medium such as a flexible disk, a magneto-optical disk, a ROM, and a CD-ROM, and a storage device such as a hard disk built in a computer system. Further, the “computer-readable recording medium” refers to a volatile memory (RAM) inside a computer system serving as a server or a client when a program is transmitted through a network such as the Internet or a communication line such as a telephone line. In addition, programs that hold programs for a certain period of time are also included.
[0050]
Further, the above program may be transmitted from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or by a transmission wave in the transmission medium. Here, the “transmission medium” for transmitting a program refers to a medium having a function of transmitting information, such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line. Further, the program may be for realizing a part of the functions described above. Furthermore, what can implement | achieve the function mentioned above in combination with the program already recorded on the computer system, and what is called a difference file (difference program) may be sufficient.
[0051]
【The invention's effect】
As described above, according to the present invention, the configuration of the reference vector table and the vector quantizer reduces the number of dimensions of data to be subjected to vector quantization from the pixel area to the block length. That is, it is possible to provide an image compression apparatus that can reduce the number of reference vectors to be compared and reduce the amount of comparison per vector. This device reproduces the direction of the largest change by vector quantization, the amount of change is smaller than that, and the small change is realized by the second quantization. Can be assigned to vector quantization, and the finer portion can be assigned to other encoding methods. As a result, it is possible to obtain an effect that the compression can be performed without reducing the image quality as compared with the case where the number of reference vectors in the vector table is blindly reduced for speeding up.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a configuration of an embodiment of the present invention.
FIG. 2 is an explanatory diagram illustrating an example of block data.
FIG. 3 is an explanatory diagram showing a maximum change direction.
FIG. 4 is an explanatory diagram showing a method of calculating the amount of change in relief.
FIG. 5 is an explanatory diagram showing a method of area division.
FIG. 6 is an explanatory diagram showing an example of an undulation change amount.
FIG. 7 is an explanatory diagram illustrating an example of a reference vector table.
FIG. 8 is an explanatory diagram illustrating an example of an intermediate calculation value.
FIG. 9 is an explanatory diagram illustrating an example of a quantization pattern.
FIG. 10 is a block diagram showing a modification of the device.
FIG. 11 is an explanatory diagram showing an example of plane data to be restored.
FIG. 12 is an explanatory diagram illustrating an example of a correction amount block.
FIG. 13 is an explanatory diagram showing an example of a decoding result.
FIG. 14 is an explanatory diagram for confirming a restoration result.
FIG. 15 is a block diagram illustrating a configuration according to another embodiment.
FIG. 16 is a block diagram illustrating a configuration according to another embodiment.
FIG. 17 is an explanatory diagram for describing effects.
[Explanation of symbols]
1 ... Block generator
2. Maximum change direction detector
20 ... Approximate data generator
3 ... Vector quantizer
4 ・ ・ ・ Cross-section shape restorer
5 Surface data restorer
6 ... Error calculator
7 ... Block encoder
71 ... BTC encoder
72 ··· Output judgment device
8 ... Entropy encoder
9 Multiplexer
11 ・ ・ ・ Block generator
12 ... vector quantizer
13 ... vector reconstructor
14 ・ ・ ・ Error calculator
15 ... Block encoder
16 ... Entropy encoder
17 ・ ・ ・ Multiplexer
21 ... Block generator
22 ... Sub-sampler
23 ... vector quantizer
24 ・ ・ ・ Vector restorer
25 ... Enlarged block generator
26 ... Error calculator
27 ... Block encoder
28 ・ ・ ・ Entropy encoder
29 ・ ・ ・ Multiplexer

Claims (15)

Block dividing means for dividing the entire image into blocks of n × n (n is a natural number of 2 or more) pixels;
a maximum change direction detecting means for calculating a change amount of luminance in the block in each of m (m is a natural number of 2 or more) directions and selecting a direction in which the change amount is the largest;
Approximation data generating means for approximating a two-dimensional luminance change of a pixel in the block to a one-dimensional luminance change only in the direction output by the maximum change direction detection means,
The approximate one-dimensional data is defined as a vector having pixel values for each shift from the block center as an element, and a vector closest to the shape of the pixel value change is searched for from a set of reference vectors prepared in advance, and this vector is identified. Vector quantization means for outputting a vector number;
Vector restoration means for restoring a vector that is the original approximate data from the vector number,
Surface data duplication means for generating restoration data in a direction perpendicular to the change direction from vector data which is a one-dimensional pixel value change in the maximum change direction, and as approximate surface data,
Error value calculation means for calculating an error value by subtracting the approximate surface data from the original image in the block, and calculating an error value of each pixel in the block,
An image compression apparatus comprising: a block encoding unit that encodes an error value distributed in the block. The encoding apparatus according to claim 1, further comprising encoding means for entropy encoding the maximum change direction output from the maximum change direction detection means, the output of the vector quantization means, and the output of the block coding means. Image compression device. The block encoding means includes:
BTC encoding means for encoding the surface distribution of the error by obtaining a value for each block that has calculated the quantitative tendency of the undulation of the error and a value for each pixel representing the distribution of the undulation,
The output of the BTC encoding means is output as it is when the quantitative tendency of the error is large in response to the output of the BTC encoding means, and the code output indicating that there is no BTC code data is output if the undulation is small. 2. The image compression device according to claim 1, further comprising a determination unit. Block dividing means for dividing the entire image into blocks of n × m (n, m is a natural number of 2 or more) pixels;
The original image in the block is set to an n × m-dimensional vector, a vector closest to the original image vector is searched for from a set of reference vectors previously expressed only with a rough expression vector, and a vector for outputting a vector number indicating the vector is output. Quantization means;
Vector restoration means for restoring a vector that is the original approximate data from the vector number,
Error calculation means for obtaining an error value by subtracting the approximate data from the original image in the block, to generate an error block,
Vector encoding means for encoding the error block,
An image compression apparatus comprising: a vector quantization unit; and an encoding unit that losslessly encodes an output of the vector encoding unit. Block dividing means for dividing the entire image into blocks of n × m (n, m is a natural number of 2 or more) pixels;
Sub-sampling means for generating a reduced image of 1 / d times and 1 / d times (d is an integer of 2 or more) from the image;
A vector quantum for outputting a vector number for identifying a vector closest to the reduced image vector from a set of reference vectors prepared in advance as a reduced image vector having each pixel of a block obtained by cutting the reduced image as an element, and outputting a vector number for identifying the vector Means,
Vector restoration means for restoring a vector that is the original approximate data from the vector number,
Enlarged data generating means for expanding the approximate vector obtained from the vector restoring means to the original size,
Error calculation means for obtaining an error value by subtracting and subtracting an approximate vector from the original image in the block, and generating an error block,
Vector encoding means for encoding the error block,
An image compression apparatus comprising: an encoding unit configured to reversibly encode an output of the vector encoding unit and the quantization unit. A block division process of dividing the entire image into blocks of n × n (n is a natural number of 2 or more) pixels;
a maximum change direction detecting step of calculating a change amount of luminance in the block in each of m (m is a natural number of 2 or more) directions and selecting a direction having the largest change amount;
An approximate data generation step of approximating a two-dimensional luminance change of a pixel in the block to a one-dimensional luminance change only in a direction obtained in the maximum change direction detection step;
The approximate one-dimensional data is defined as a vector having pixel values for each shift from the block center as an element, and a vector closest to the shape of the pixel value change is searched for from a set of reference vectors prepared in advance, and this vector is identified. A vector quantization process for outputting vector numbers,
A vector restoration process of restoring a vector that is the original approximate data from the vector number,
Surface data restoration process of generating restoration data in a direction perpendicular to the change direction from vector data that is a one-dimensional pixel value change in the maximum change direction and setting it as approximate surface data,
An error value calculating step of calculating an error value by subtracting the approximate surface data from the original image in the block, and calculating an error value of each pixel in the block,
A block encoding step of encoding an error value distributed in the block. 7. The image compression apparatus according to claim 6, further comprising a coding step of entropy coding the maximum change direction obtained in the maximum change direction detection step, the output of the vector quantization step, and the output of the block coding step. Method. The block encoding process includes:
A BTC encoding process of encoding the surface distribution of the error by obtaining a value for each block that has calculated the quantitative tendency of the undulation of the error and a value for each pixel representing the distribution of the undulation;
When the output of the BTC encoding process receives the output and the error has a large quantitative tendency, the output of the BTC encoding process is output as it is, and when the amount of undulation is small, a code output indicating no BTC code data is output. 7. The image compression method according to claim 6, comprising a determining step. A block division process of dividing the entire image into blocks of n × m (n, m is a natural number of 2 or more) pixels;
The original image in the block is set to an n × m-dimensional vector, a vector closest to the original image vector is searched for from a set of reference vectors previously expressed only with a rough expression vector, and a vector for outputting a vector number indicating the vector is output. The quantization process,
A vector restoration process of restoring a vector that is the original approximate data from the vector number,
Calculating an error value by subtracting the approximate data from the original image in the block, an error calculation step of generating an error block,
A vector encoding process for encoding the error block;
An image compression method, comprising: the vector quantization step; and an encoding step of losslessly encoding an output of the vector encoding step. A block division process of dividing the entire image into blocks of n × m (n, m is a natural number of 2 or more) pixels;
A sub-sampling process of generating a reduced image of 1 / d times and 1 / d times (d is an integer of 2 or more) from the image;
A vector quantum for outputting a vector number for identifying a vector closest to the reduced image vector from a set of reference vectors prepared in advance as a reduced image vector having each pixel of a block obtained by cutting the reduced image as an element, and outputting a vector number for identifying the vector Process and
A vector restoration process of restoring a vector that is the original approximate data from the vector number,
An enlarged data generation step of expanding the approximate vector obtained from the vector restoration step to the original size,
Calculating an error value by subtracting and subtracting an approximate vector from an original image in the block, and an error calculation process of generating an error block;
A vector encoding process for encoding the error block;
An image compression method, comprising: an encoding step of reversibly encoding an output of the vector encoding step and the quantization step. A block division process of dividing the entire image into blocks of n × n (n is a natural number of 2 or more) pixels;
a maximum change direction detection process of calculating a change amount of luminance in the block in each of m (m is a natural number of 2 or more) directions and selecting a direction having the largest change amount;
Approximate data generation processing that approximates a two-dimensional luminance change of a pixel in the block to a one-dimensional luminance change only in the direction output by the maximum change direction detection processing;
The approximate one-dimensional data is defined as a vector having pixel values for each deviation from the block center as an element, and a vector closest to the shape of the pixel breaking change is searched for from a set of reference vectors prepared in advance, and this vector is identified. Vector quantization processing for outputting vector numbers,
Vector restoration processing for restoring a vector that is original approximate data from the vector number,
Surface data restoration processing that generates restoration data in a direction perpendicular to the change direction from vector data that is a one-dimensional pixel value change in the maximum change direction, and sets it as approximate surface data;
Calculating an error value by subtracting the approximate surface data from the original image in the block, and calculating an error value of each pixel in the block;
An image compression program for causing a computer to perform a block encoding process for encoding an error value distributed in the block. 12. The computer according to claim 11, further comprising an encoding process for entropy encoding the maximum change direction obtained in the maximum change direction detection process, the output of the vector quantization process, and the output of the block encoding process. Image compression program. The block encoding process includes:
A BTC encoding process for encoding a surface distribution of the error by obtaining a value for each block that calculates a quantitative tendency of the undulation of the error and a value for each pixel representing the distribution of the undulation;
The output of the BTC encoding process is output as it is when the quantitative tendency of the error is large in response to the output of the BTC encoding process, and the code output indicating no BTC code data is output when the undulation is small. 12. The image compression program according to claim 11, comprising a determination process. A block division process of dividing the entire image into blocks of n × m (n, m is a natural number of 2 or more) pixels;
The original image in the block is set to an n × m-dimensional vector, a vector closest to the original image vector is searched for from a set of reference vectors previously expressed only with a rough expression vector, and a vector for outputting a vector number indicating the vector is output. Quantization processing,
Vector restoration processing for restoring a vector that is original approximate data from the vector number,
Calculating an error value by subtracting the approximate data from the original image in the block, an error calculation process for generating an error block,
A vector encoding process for encoding the error block;
An image compression program for causing a computer to perform a vector quantization process and an encoding process of losslessly encoding an output of the vector encoding process. A block division process of dividing the entire image into blocks of n × m (n, m is a natural number of 2 or more) pixels;
Sub-sampling processing for generating a reduced image of 1 / d times and 1 / d times (d is an integer of 2 or more) from the image;
A vector quantum for outputting a vector number for identifying a vector closest to the reduced image vector from a set of reference vectors prepared in advance as a reduced image vector having each pixel of a block obtained by cutting the reduced image as an element, and outputting a vector number for identifying the vector Conversion processing,
Vector restoration processing for restoring a vector that is original approximate data from the vector number,
Enlarged data generation processing for expanding the approximate vector obtained from the vector restoration processing to the original size;
Calculating an error value by subtracting and subtracting an approximate vector from the original image in the block, an error calculation process of generating an error block;
A vector encoding process for encoding the error block;
An image compression program for causing a computer to perform a vector encoding process and an encoding process of losslessly encoding an output of the quantization process.

Images