JP2011249954A - Ac component prediction system and ac component prediction program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an AC component prediction system and an AC component prediction program which can further improve prediction accuracy in AC component prediction.SOLUTION: A low-frequency AC component prediction part 31 predicts the low-frequency AC component of an object block by AC component prediction, and a high-frequency AC component prediction part 32 predicts the high-frequency AC component of the object block by weighting prediction. A prediction processing part 33 adds the low-frequency AC component and the high-frequency AC component in order to calculate the prediction value of the AC component of the object block. An AC component restoration part 35 adds the prediction value of the AC component and the residual component from the original image in order to calculate the restoration value of the AC component of the object block. A high-frequency AC component calculation part 38 calculates the high-frequency AC component used as information of the processed block from the next processing, based on the difference between the restoration value of the AC component and the low-frequency AC component.

Description

  The present invention relates to an AC component prediction system and an AC component prediction program.

  Conventionally, image processing called AC component prediction (ACP) is known. As illustrated in FIG. 1, the AC component prediction refers to information on a peripheral area (for example, PBl, PBr, PBt, PBb) of a target block to be processed, and subblocks obtained by subdividing the block PBs. A method for obtaining information c00 to c11. In the alternating current component prediction, the spatial resolution between the target block PBs and the surrounding area located in the vicinity thereof is utilized to maintain the DC value (direct current component) of the target block PBs, and the spatial resolution is further improved. A high AC value (AC component) is calculated. Patent Documents 1 to 3 disclose recursive AC component predictive coding (RACP) that sequentially subdivides a target block and hierarchically encodes image data. Specifically, first, the predicted pixel values of sub-blocks obtained by subdividing the target block are sequentially calculated while sequentially shifting the target block in a predetermined direction on the image plane. Next, a difference between the predicted pixel value and the original pixel value (true value) is calculated as a residual component. The residual component is subjected to irreversible transformation and entropy coding to generate an alternating current component of an image that is part of the compressed image data. The above processing is performed by, for example, an 8 × 8 pixel block (top layer), a 4 × 4 pixel block (second layer), a 2 × 2 pixel block (third layer), and 1 × 1. It is recursively repeated in a hierarchical structure consisting of pixel blocks (lowest hierarchy).

Japanese Patent No. 4000157 Japanese Patent No. 3774201 Japanese Patent No. 3700976

  By the way, as is well known, a block on the image plane (real space) can theoretically be represented by a DC component DC of this block and three Hadamard bases orthogonal to each other by mapping to the Hadamard space. As shown in FIG. 2, the block composed of 2 × 2 pixels has a DC component DC of this block and a scalar, no matter what the four pixel values c00 to c11 exist in the block. It can be expressed as the sum of three multiplied Hadamard bases. Here, in the first Hadamard basis (α basis) following the average value DC, +1 and −1 are arranged in the horizontal direction. In the second Hadamard basis (β basis) that follows this, +1 and −1 are arranged in the vertical direction. Then, in the third Hadamard base (γ base) following this, +1 and −1 are diagonally crossed and arranged. These Hadamard bases are orthogonal to each other, as is clear from the fact that an arbitrary product is selected and the inner product is always 0. The multiplier α is an α-based scalar value (α component), the multiplier β is a β-based scalar value (β component), and the multiplier γ is a γ-based scalar value (γ component). In the mapping from the real space to the Hadamard space, a well-known Hadamard transform is used, whereby the α component, β component and γ component are uniquely specified from the pixel values c00 to c11. On the other hand, in the mapping from the frequency space to the real space, the well-known inverse Hadamard transform is used, whereby the pixel values c00 to c11 are uniquely specified from the α component, β component, and γ component. This means that what is equivalent to processing in real space may be performed in Hadamard space. The same applies to the AC component prediction, and predicting the pixel values c00 to c11 in the block is equivalent to predicting the α component, the β component, and the γ component on the assumption that the average value DC of the block is known. It is. It should be noted that the pixel values c00 to c11 can be satisfactorily approximated with only the α component and the β component (even if the γ component is regarded as 0).

  It is preferable that the image prediction accuracy by the AC component prediction is high. This is because the higher the prediction accuracy, the higher the correlation between the predicted image as a set of predicted values and the original image, and the smaller the prediction component corresponding to the difference between the two. The reduction of the residual component means that a tendency that the appearance frequency of the residual component viewed statistically tends to be close to 0 increases. Therefore, for example, when entropy coding is performed on the residual component to compress the data, the compression rate can be increased by reducing the residual component. In AC component prediction, the closer the distance from the target block is, that is, the higher the spatial correlation, the smaller the divisor of the change amount (slope) of the pixel value, and the more accurate prediction becomes possible. In this regard, in Patent Documents 1 to 3, the residual component in the previous processing (that is, the prediction error with respect to the original image generated in the previous processing) is used as the predicted value of the AC component to be calculated in the current processing. Without any consideration, the prediction of each block is performed independently of each other, leaving room for improvement in prediction accuracy.

  Therefore, an object of the present invention is to further improve the prediction accuracy in AC component prediction.

  In order to solve such a problem, the first invention includes a low-frequency AC component prediction unit, a high-frequency AC component prediction unit, a prediction processing unit, an AC component restoration unit, and a high-frequency AC component calculation unit. Provided is an AC component prediction system that performs AC component prediction in units of blocks by iterative processing for each block set on a plane. The low-frequency AC component prediction unit predicts the low-frequency AC component of the target block by AC component prediction using the DC component of the block corresponding to the reference area of the target block to be processed in the current process. The high-frequency AC component prediction unit predicts the high-frequency AC component of the target block by weighted prediction using the high-frequency AC component of the processed block. This processed block is located in the vicinity of the target block and has already been processed in the previous process. The prediction processing unit calculates the predicted value of the AC component of the target block by adding the low-frequency AC component predicted by the low-frequency AC component prediction unit and the high-frequency AC component predicted by the high-frequency AC component prediction unit. To do. The AC component restoration unit calculates the restoration value of the AC component of the target block by adding the prediction value calculated by the prediction processing unit and the residual component of the original image. The high-frequency AC component calculation unit is information on blocks that have been processed in the next and subsequent processes based on the difference between the restoration value calculated by the AC component restoration unit and the low-frequency AC component predicted by the low-frequency AC component prediction unit. The high frequency alternating current component used as is calculated.

  Here, in the first invention, the high-frequency alternating current component prediction unit responds to the prediction direction of the processed block with reference to the target block so that the error of the pixel value between the target block and the processed block becomes small. Thus, it is preferable to perform the weighted prediction after inverting the sign of the high-frequency AC component of the processed block. Here, the AC component preferably includes a horizontal component that defines a component of the Hadamard base in the horizontal direction. In this case, the high-frequency AC component prediction unit adds the horizontal component related to the processed block close to the target block in the vertical direction and the value obtained by inverting the sign of the horizontal component related to the processed block close to the target block in the horizontal direction. By doing so, it is desirable to perform the above weighted prediction regarding the high-frequency AC component included in the horizontal component. The alternating current component preferably includes a vertical component that defines a Hadamard basis component in the vertical direction. In this case, the high-frequency AC component prediction unit adds the value obtained by inverting the sign of the vertical component related to the processed block close to the target block in the vertical direction and the vertical component related to the processed block close to the target block in the horizontal direction. By doing so, it is desirable to perform the weighted prediction regarding the high-frequency AC component included in the vertical component.

  The AC component prediction system according to the first invention may be applied to an encoder that performs irreversible compression of an image. In this case, it is preferable that the AC component restoration unit includes a subtracter, an irreversible conversion unit, an entropy encoding unit, an inverse conversion unit, and an adder. The subtractor calculates a difference between the predicted value calculated by the prediction processing unit and the true value in the original image as a residual component. The irreversible conversion unit performs irreversible conversion on the residual component calculated by the subtractor. The entropy encoding unit performs entropy encoding on the residual component output from the lossy conversion unit, and outputs compressed image data. The inverse transform unit restores the residual component by performing a process opposite to the irreversible transform on the residual component output from the irreversible transform unit. The adder adds the residual component output from the inverse transform unit to the prediction value calculated by the prediction processing unit.

  The AC component prediction system according to the first invention may be applied to a decoder that expands an irreversibly compressed image. In this case, the AC component restoration unit includes an inverse conversion unit and an adder. The inverse transform unit restores a residual component from the original image by performing inverse processing of irreversible transformation and entropy encoding performed at the time of image compression on the compressed data of the irreversibly compressed image. The adder outputs a decompressed image by adding the residual component output from the inverse transform unit to the prediction value calculated by the prediction processing unit.

  On the other hand, 2nd invention provides the alternating current component prediction program which makes a computer perform alternating current component prediction performed per block by the repeating process for every block set on the image plane. This program includes a first step of predicting a low-frequency AC component of a target block by an AC component prediction using a DC component of a block corresponding to a reference area of the target block to be processed in the current process, and the target block A second step of predicting the high-frequency alternating current component of the target block by weighted prediction using the high-frequency alternating current component of the processed block already processed in the previous processing, and a low-frequency alternating current component, The third step of calculating the predicted value of the alternating current component of the target block by adding the high frequency alternating current component and the alternating current of the target block by adding the predicted value and the residual component of the original image Based on the difference between the restored value and the low-frequency alternating current component in the fourth step of calculating the restored value of the component, High-frequency AC component used Te to execute a fifth step of calculating the computer.

  Here, in the second invention, the second step corresponds to the prediction direction of the processed block with reference to the target block so that the error of the pixel value between the target block and the processed block becomes small. Thus, the step of performing weighted prediction after inverting the sign of the high-frequency AC component of the processed block may be used. Here, the AC component preferably includes a horizontal component that defines a component of the Hadamard base in the horizontal direction. In this case, the second step adds the horizontal component related to the processed block close to the target block in the vertical direction and the value obtained by inverting the sign of the horizontal component related to the processed block close to the target block in the horizontal direction. By doing so, it is desirable to be a step of performing weighted prediction regarding the high-frequency AC component included in the horizontal component. The alternating current component preferably includes a vertical component that defines a Hadamard basis component in the vertical direction. In this case, the second step adds the value obtained by inverting the sign of the vertical component related to the processed block close to the target block in the vertical direction and the vertical component related to the processed block close to the target block in the horizontal direction. By doing so, it is desirable to be a step of performing weighted prediction regarding the high-frequency AC component included in the vertical component.

  Further, the AC component prediction program according to the second invention may be applied to an encoding program for irreversibly compressing an image. In this case, the fourth step includes a step of calculating a difference between the predicted value and the true value in the original image as a residual component, a step of performing irreversible transformation on the residual component, and an entropy code on the residual component. And outputting the compressed data of the image, performing a process reverse to the lossy conversion on the residual component, restoring the residual component, and adding the residual component to the predicted value It is desirable to have steps.

  Furthermore, the AC component prediction program according to the second invention may be applied to a decoding program that expands an irreversibly compressed image. In this case, in the fourth step, the residual from the original image is obtained by performing reverse processing of the lossy transformation and entropy coding performed at the time of image compression on the compressed data of the lossy compressed image. It is desirable to have a step of restoring the component and a step of outputting the decompressed image by adding the residual component to the predicted value.

  According to the first and second aspects of the invention, the predicted value of the AC component related to the target block is corrected by adding the high-frequency AC component to this instead of the low-frequency AC component predicted by the AC component prediction. A value is used. The high-frequency alternating current component to be added in the current prediction is calculated by weighted prediction using the high-frequency alternating current component of the processed block, that is, the value already calculated in the previous processing. The high-frequency AC component corresponds to a restoration value of the AC component obtained by adding the residual component to the predicted value (including both the low-frequency AC component and the high-frequency AC component) and excluding the low-frequency AC component. Therefore, the magnitude of the high-frequency alternating current component depends on the residual component, and the higher the alternating current component becomes, the larger the residual component with the original image becomes. By predicting the current high-frequency alternating current component from the previous high-frequency alternating current component and adding this to the current low-frequency alternating current component, the previous residual component according to the current predicted value is reflected. In addition, the current high frequency alternating current component is used to predict the high frequency alternating current component from the next time, and this is added to the low frequency alternating current component from the next time on, so that the current residual component is reflected in the predicted value from the next time. . By repeating such processing, the residual component of a certain block is sequentially propagated and reflected on the other blocks. As a result, since a component that cannot be captured only by the AC component prediction is added to the predicted value, the tendency of the predicted image to approach the original image is increased, and the prediction accuracy can be further improved.

Explanatory drawing of sequential processing of blocks on image plane Diagram of relationship between block on image plane and Hadamard space representation Overall configuration of RACP encoder Operation timing chart in parallel processing Operation timing chart in sequential processing Configuration diagram of hierarchical processing unit in RACP encoder Explanatory diagram of positional relationship between block center position and nearby prediction points Explanation of calculation of neighboring pixel values using three points (λ = 1/2) Explanation of calculation of neighboring pixel values using three points (λ = 1/4) Explanation of calculation of neighboring pixel values using three points (λ = 1/8) Explanation of calculation of neighboring pixel values using two points (λ = 1/2) Explanatory drawing of sign inversion in weighted prediction of horizontal component (αhigh) Explanatory drawing of sign inversion in weighted prediction of vertical component (βhigh) Illustration of residual propagation starting from a certain block RACP encoding program flowchart Overall configuration of RACP decoder Configuration diagram of hierarchical processing unit in RACP decoder RACP decoding program flowchart

  Hereinafter, as an embodiment to which the AC component prediction according to the present invention is applied, recursive AC component prediction coding (RACP: Recursive ACP) will be described as an example.

(RACP encoder)
FIG. 3 is an overall configuration diagram of the RACP encoder. This encoder has a DC calculation unit 1, a DC encoding unit 2, and hierarchical processing units 3a to 3c, and is configured by four layers in this embodiment. Data TDCn, DCn (n = 0, 1, 2) output from these units 1, 2, 3a to 3c are temporarily stored in a buffer (storage unit) not shown.

  The DC calculation unit 1 divides an input image to be compressed into blocks of a preset size, and outputs the DC component of each block, that is, the average value of pixels included in the block, as TDC0 to TDC3. Here, TDC0 is a DC component of an 8 × 8 pixel block (hereinafter referred to as “8 × 8 block”), TDC1 is a DC component of a 4 × 4 pixel block (hereinafter referred to as “4 × 4 block”), and TDC2 is 2 It is a DC component of a block of × 2 pixels (hereinafter referred to as “2 × 2 block”). TDC3 is a direct current component of a 1 × 1 pixel block (hereinafter referred to as “1 × 1 block”), that is, a pixel value itself of a pixel which is a minimum unit of an image. The direct current components TDC0 to TDC3 (true values) calculated directly from the input image are distinguished from the direct current components DC0 to DC2 (restored values) restored through processing in the units 2 and 3a to 3c. Please note that. Since both values are accompanied by irreversible transformation in the process of encoding, they do not exactly match.

  The DC encoding unit 2 and the hierarchical processing units 3a to 3c include 8 × 8 blocks (top layer), 4 × 4 blocks (second layer), 2 × 2 blocks (third layer), and 1 × 1 block. In the four-layer structure composed of (the lowest layer), the layer processing assigned to itself is performed.

  The DC encoding unit 2 in the highest layer performs differential pulse code modulation (DPCM) and entropy encoding on the DC component TDC0 for each 8 × 8 block read from the buffer. Difference pulse code modulation encodes the difference between the DC components TDC0 for blocks adjacent to each other on the image plane. Then, entropy coding such as Huffman coding or arithmetic coding is performed on the coded difference after quantization. The data that has undergone such processing is output as a DC component DC0 of an image that is a part of the compressed data, and an 8 × 8 block DC component is supplied to the hierarchical processing unit 3a of the immediately lower hierarchy. A restored (inverse quantized) value DC0 is output.

  The three hierarchical processing units 3a to 3c are configured to use the direct current component TDCn (n = 1, 2, 3, and so on) generated by the DC calculation unit 1 and the direct current component DCn-1 generated by the immediately higher hierarchy. The AC component ACn of the image as a part of the compressed data is output by the processing including the input AC component prediction. At the same time, the hierarchy processing units 3a to 3b (excluding 3c) restore the DC component DCn from the AC component ACn. The restored DC components DC0, 1, 2 are output for supply to the immediately lower hierarchy.

  Specifically, the hierarchical processing unit 3a in the second hierarchical level undergoes processing such as AC component prediction and residual component calculation for 8 × 8 blocks, and 4 × 4 blocks of AC components AC1 and 4 Outputs the DC component DC1 of x4 blocks. In this hierarchy, the DC component DC0 of 8 × 8 blocks generated in the highest hierarchy is used as information to be referred to in the AC component prediction. At the same time, in order to calculate the residual component, the DC component TDC1 of 4 × 4 blocks calculated by the DC calculation unit 1 is input. The hierarchy processing unit 3b in the third hierarchy outputs 2 × 2 blocks of AC components AC2 and 2 × 2 blocks of DC components DC2 through processing such as AC component prediction for 4 × 4 blocks. To do. In this hierarchy, the DC component DC1 of 4 × 4 blocks generated in the second hierarchy is used as reference information for AC component prediction. At the same time, in order to calculate a residual component, a DC component TDC2 of 2 × 2 blocks calculated by the DC calculation unit 1 is input. The hierarchy processing unit 3c in the lowest hierarchy outputs an AC component AC3 of 1 × 1 block through processing such as AC component prediction for 2 × 2 blocks (calculation of DC3 is not necessary). In this hierarchy, the DC component DC2 of 2 × 2 blocks generated in the third hierarchy is used as reference information for AC component prediction. At the same time, in order to calculate the residual component, the DC component TDC3 of 1 × 1 block calculated by the DC calculation unit 1 is used.

  As described above, in the hierarchical structure in which the sizes of the blocks to be processed are different, the DC encoding unit 2 and the hierarchical processing units 3a to 3c are linked to each other so that image processing mainly based on AC component prediction is recursively executed. Is done. As a result, the DC component DC0 of the image and the AC components AC1 to AC3 are output as compressed data. The compressed data includes accompanying information such as a Huffman table in addition to these.

  As shown in FIG. 1, a plurality of blocks serving as processing units for AC component prediction are set on an image plane by dividing an image to be compressed (one frame image or a partial image thereof) vertically and horizontally. The block size is set so as to decrease step by step as the hierarchy becomes lower. The size of the block to be processed in a certain hierarchy matches the size of the sub-block subdivided in the hierarchy located immediately above it. In other words, the size of the upper-layer sub-block matches the size of the lower-layer block located immediately below it. Processing of the entire image in a certain hierarchy is achieved by repeating the processing while sequentially shifting the blocks to be processed on the screen and processing all the blocks in the image. As shown in the figure, the shift direction (scan direction) may be a line-sequential scan along the horizontal direction, but may be arbitrarily set including the one along the vertical direction. Further, the shift directions in the respective hierarchies are not necessarily the same. In the case of the illustrated shift direction, if the block to be processed is “PBs”, the upper side of the horizontal line to which the target block PBs belongs is the processed region, and the lower side is the unprocessed region. As for the horizontal line of the target block PBs, the left side of the target block PBs is a processed area, and the right side thereof is an unprocessed area. For example, eight blocks PBt, PBtt, PBl, PBll, PBb, PBbb, PBr, and PBrr having the same block size are used as reference areas that are referred to in the processing of the target block PBs. Among these reference blocks, the blocks PBt, PBl, PBb, and PBr are particularly referred to as “inner blocks”, and the blocks PBtt, PBll, PBbb, and PBrr are referred to as “outer blocks”. The inner block PBt is adjacent to the upper side of the target block PBs, and the outer block PBtt is adjacent to the outer side of the inner block PBt in the upper prediction direction connecting the inner block PBt and the target block PBs. In the same relationship, the blocks PBl and PBll are adjacent to each other in the left prediction direction, the blocks PBb and PBbb are adjacent to each other in the lower prediction direction, and the blocks PBr and PBrr are adjacent to each other in the right prediction direction. By the AC component prediction of the target block PBs, the average pixel value (DC component of the sub block) of each of the sub blocks divided into four is calculated. Hereinafter, when subblocks are subdivided into blocks, the upper left subblock is identified by the subscript “00”, the upper right is “01”, the lower right is “11”, and the lower left is “10”.

  The processing of the RACP encoder may be either parallel processing or sequential processing. FIG. 4 is an operation timing chart in the parallel processing. First, the DC calculation unit 1 operates to generate DC components TDC0 to TDC3. The DC encoding unit 2 and the hierarchical processing units 3a to 3c operate in parallel with the completion of the generation of all the direct current components TDC0 to TDC3. However, the timing at which these operations start is not the same, and the lower the hierarchy, the later the start timing. This delay is caused by processing delay due to sequential shift of the upper layer. Referring to FIG. 1, when processing a block PBs in a certain hierarchy, the condition (hierarchical process start condition) is that the processing of the reference blocks PBt to PBrr is completed in the immediately higher hierarchy. In other words, in the upper hierarchy, the process of the block PBs in the lower hierarchy cannot be started unless the sequential shift proceeds and the process up to the block PBbb is finished. This is the reason why the operation delay occurs in the lower layer. However, the pipeline parallel processing shown in FIG. 4 is faster than the sequential processing shown in FIG. On the other hand, in the sequential processing of FIG. 5, the processing of the immediately lower layer is started when all processing of a certain layer is completed. Accordingly, the hierarchical processing start condition is naturally satisfied even in the sequential processing.

  FIG. 6 is a configuration diagram of the hierarchy processing unit 3 (generic name of 3a to 3c) in the RACP encoder. The hierarchical processing units 3 in the respective layers have almost the same basic configuration and operation except that the handled block sizes are different (the lowest hierarchical processing unit 3c can be partially simplified). The hierarchical processing unit 3 includes a low-frequency AC component prediction unit 31, a high-frequency AC component prediction unit 32, a prediction processing unit 33, a Hadamard transform unit 34, an AC component restoration unit 35, an entropy encoding unit 36, and an inverse The Hadamard transform unit 37 and the high-frequency AC component calculation unit 38 are mainly configured.

  The low frequency AC component prediction unit 31 is supplied with information necessary for predicting the low frequency AC component of the target block PB after being read from the buffer. In the present embodiment, as information necessary for this processing, as shown in FIG. 1, the DC component DC (n-1) of the target block PBs to be processed this time and the eight reference blocks positioned around it. A DC component DC (n-1) of PBt to PBrr (reference region) is used. The DC component DC (n−1) of the reference blocks PBt to PBrr is independent of whether the block has been processed or not processed, in other words, whether or not the information of the sub-blocks in the block is calculated. Regardless, the DC component DC (n-1) calculated in the immediately higher hierarchy (n-1) is uniformly used.

  The low-frequency AC component prediction unit 31 predicts the low-frequency AC component AClow of the target block PBs by AC component prediction using the DC component DC (n-1) of the reference blocks PBt to PBrr. At that time, for each of the prediction directions of the top, bottom, left, and right, a neighborhood prediction point is set at a position closer to the target block PBs than the center position of the inner block, and a neighborhood pixel value at this neighborhood prediction point is predicted and calculated. The AC component prediction is performed using the neighboring pixel values calculated individually for each prediction direction.

  The low-frequency AC component prediction unit 31 performs the AC component prediction using the DC component DC (n-1) of the reference blocks PBt to PBrr corresponding to the reference region of the target block PBs, thereby performing the low-frequency AC component AClow = of the target block PBs. Predict {αlow, βlow, γlow}. This prediction is performed based on the following formulas 1 and 2. Hereinafter, the DC component value DC (n−1) of a certain block (for example, “PBr”) is described by using the suffix (for example, the suffix “R”) of the block.

(Formula 1)
c00 = S + (αlow + βlow + γlow)
c01 = S + (− αlow + βlow−γlow)
c10 = S + (αlow−βlow−γlow)
c11 = S + (− αlow−βlow + γlow)

(Formula 2)
αlow = (L'-R ') / 6
βlow = (T′−B ′) / 6
γlow = 0
L ′ = L + (S−LL) / 8
R ′ = R + (S−RR) / 8
T ′ = T + (S−TT) / 8
B ′ = B + (S−BB) / 8

  Here, c00 to c11 are predicted pixel values of sub-blocks s00 to s11 obtained by subdividing the target block PBs. Predicting the pixel values c00 to c11 of the real space from the reversibility of the mapping between the real space and the Hadamard space is based on the assumption that the DC component S of the target block PBs is known, and the low frequency of the Hadamard space. This is equivalent to predicting the AC component AClow, that is, αlow, βlow, and γlow. However, it is well known from the prior art documents exemplified as the prior art that the pixel values c00 to c11 can be satisfactorily approximated by omitting the γ component that is difficult to linearly predict and using only the α and β components. αlow is a low-frequency AC component related to the horizontal component that defines the horizontal Hadamard basis component, and βlow is a low-frequency AC component related to the vertical component that defines the vertical Hadamard basis component. Γlow is a low-frequency alternating current component related to the other components orthogonal to these two bases. Formula 2 is a formula for calculating the neighboring pixel values L ′, R ′, T ′, and L ′ when λ = ½. Regarding λ, in the direction from the target block PBs to the reference block, the block boundary between the target block and the inner block is set to λ = 0, and the distance from there to the center position of the inner block is set to λ = 1. The neighboring pixel values L ′, R ′, T ′, and B ′ are values predicted by linear prediction or the like for the neighboring prediction points that are closer to the target block PBs than the center position of the inner block.

  FIG. 7 is an explanatory diagram of the positional relationship between the block center position and the vicinity prediction point at λ = 1/2. The central pixel value at the center position of each block is regarded as the DC component DC (n−1) of the block itself. In order to match the scale of λ, the length of one side of the square when the block is divided into four is set to 1, and the length of one side of the square when the sub-block is further divided into four is set to ½. In this case, the center-to-center distance between adjacent blocks is 2, and the distance between the center position of the target block PBs and each of the neighboring prediction points (its neighboring pixel values = L ′, R ′, T ′, B ′) is all. 3/2. For each prediction direction, the neighborhood pixel value (eg, R ′) of the neighborhood prediction point uses at least the center pixel value (eg, R) of the inner block and the center pixel value (eg, RR) of the outer block, and Prediction / calculation is performed based on the distance between the center positions of each other and the distance between the center position of the block and the neighboring prediction points.

  FIG. 8 is an explanatory diagram of calculation of the neighboring pixel value R ′ regarding the right prediction direction using three points at λ = 1/2. First, the inclination θ is calculated by dividing the difference between the center pixel value S of the target block PBs and the center pixel value RR of the right outer block PBrr by the center-to-center distance 4 of these blocks. This inclination θ corresponds to the change rate of the pixel value per unit distance of the target block PBs and the right outer block PBrr. Next, on the assumption that the pixel values change uniformly in the same prediction direction, the distance 1 between the center position of the right inner block PBr and the right neighboring prediction point is set to the gradient θ in the right prediction direction. By multiplying by 2, a change amount Δp from the center pixel value R of the right inner block PBr is calculated. Then, by adding the change amount Δp to the central pixel value R of the right inner block PBr, the neighboring pixel value R ′ of the right neighboring prediction point is calculated. In the same manner, the neighborhood pixel values T ′, L ′, and B ′ of the neighborhood prediction points that are shared by λ = ½ at the upper side are calculated.

  From the spatial correlation of the image, the neighboring pixel values T ′ to R ′ of the neighboring prediction points close to the target block PBs are those of the target block PBs that are more than the central pixel values T to R of the inner blocks PBt to PBr farther than this. There is a tendency to reflect the characteristic deeply, and it can be considered that it is closer to the true value. Therefore, by using the neighboring pixel values T ′ to R ′ instead of the center pixel values T to R themselves, it is possible to perform AC component prediction with higher prediction accuracy. When a part of the peripheral region is missing like an edge on the image plane, the low-frequency AC component prediction unit 31 exceptionally uses the center pixel value S of the target block PBs as information on the portion. Use. In the actual processing, it is natural that the above formulas 1 and 2 may be integrated and handled as a single formula.

  The low-frequency AC component prediction unit 31 predicts (typically linear prediction) neighboring pixel values T ′ to R ′ that are closer to the target block PBs with a resolution (λ ≦ 1) that is higher than ½ of one side of the block. The AC component prediction using this is performed. Therefore, the value of λ may be set to an arbitrary value satisfying λ ≦ 1 other than ½. In particular, if λ is set to a value smaller than 1/2 (such as 1/4 or 1/8), it is possible to ensure an accuracy that is equal to or higher than the resolution of the sub-block. FIG. 9 is an explanatory diagram for calculating the neighboring pixel values T ′ to R ′ using three points at λ = 1/4. In this case, the neighboring pixel values T ′ to R ′ and the low-frequency AC component AClow are determined from a linear predictive approach that takes into account the distance between the center positions of the blocks and the distance between the center position of the blocks and the neighboring prediction points. = {Αlow, βlow, γlow} is calculated from Equation 3.

(Formula 3)
αlow = (L′−R ′) / 5
βlow = (T′−B ′) / 5
γlow = 0
L ′ = L + (S−LL) × 3/16
R ′ = R + (S−RR) × 3/16
T ′ = T + (S−TT) × 3/16
B ′ = B + (S−BB) × 3/16

  FIG. 10 is an explanatory diagram of calculation of neighboring pixel values using three points at λ = 1/8. In this case, the neighboring pixel values T ′ to R ′ and the low-frequency AC component AClow = {αlow, βlow, γlow} are calculated from Equation 4.

(Formula 4)
αlow = (L′−R ′) × 4/18
βlow = (T′−B ′) × 4/18
γlow = 0
L ′ = L + (S−LL) × 7/32
R ′ = R + (S−RR) × 7/32
T ′ = T + (S−TT) × 7/32
B ′ = B + (S−BB) × 7/32

  When the AC component prediction is performed by referring only to the reference blocks PBt to PBrr having the same size as the target block PBs as in the three cases illustrated (λ = 1/2, 1/4, 1/8), λ = It is optimal to predict a pixel at 1/4 position. Further, although it is somewhat inferior to the linear prediction using the three points described above, the neighboring pixel values T ′ to R ′ can be calculated as in Equation 5 by the linear prediction using two points as shown in FIG. Good. Taking λ = 1/2 as an example, for example, the neighboring pixel value R ′ is a change rate of pixel values per unit distance of the right inner / right outer blocks PBr and PBrr (slope θ = (R−RR) / 2). Is multiplied by ½ which is the distance between the center position of the inner block and the neighboring prediction point, and the value Δp obtained by this multiplication is added to the center pixel value R of the inner block PBr.

(Formula 5)
αlow = (L'-R ') / 6
βlow = (T′−B ′) / 6
γlow = 0
L ′ = (5L−LL) / 4
R ′ = (5R−RR) / 4
T ′ = (5T−TT) / 4
B ′ = (5B−BB) / 4

  For example, you may perform alternating current component prediction based on the following Numerical formula 6 currently disclosed by the patent 3700976. In Equation 6, the division value 8 is uniquely specified according to the distance between adjacent blocks in linear prediction.

(Formula 6)
c00 = S + (U + L-B-R) / 8
c01 = S + (U + R−B−L) / 8
c10 = S + (B + L−U−R) / 8
c11 = S + (B + R−UL) / 8

  In the method of Formula 6, only the upper, lower, left, and right reference blocks PBt to PBr having the same size as the target block PBs are used as the reference area of the target block PBs, and information on subblocks generated as needed by iterative processing for each block. (AC component) is not necessary. In addition to the method using only blocks of the same size as the target block PBs, as disclosed in Japanese Patent No. 3774201 and Japanese Patent No. 4000157, processed regions (for example, PBt and PB1 in FIG. 1) For, subblock information (AC component) obtained by subdividing a block may be used in combination. Furthermore, the reference area may be set as appropriate including not only the block or sub-block directly adjacent to the target block PBs but also a block one block away from the block or sub-block.

  The point common to all the AC component predictions exemplified here is that at least a block having the same size as the target block PBs is used as the reference region of the target block PBs. This is the low-frequency AC component according to this embodiment. Is a necessary condition. The low frequency AC component AClow calculated by the low frequency AC component prediction unit 31 is supplied to the prediction processing unit 33 and the high frequency AC component calculation unit 38.

  On the other hand, information necessary for high frequency AC component prediction of the target block PB is read from the buffer and supplied to the high frequency AC component prediction unit 32 provided in parallel with the low frequency AC component prediction unit 31. Information necessary for this processing is the high-frequency AC component AChigh (n) of the processed block that is located in the vicinity of the target block PBs and has already been processed in the previous processing. In this embodiment, the information of the target block PBs The reference blocks PBt and PBl located in the vicinity of the vertical / horizontal direction are used. The high-frequency AC component AC (t) high related to the processed block PBt in the vertical direction (upward) has already been calculated by the high-frequency AC component calculation unit 38 in the previous processing in the same hierarchy (n) using this as the target block PBs. At the time of starting this processing, it is stored in the buffer. Similarly, the high-frequency AC component AC (l) high related to the horizontal (left) block PBl has already been obtained by the high-frequency AC component calculation unit 38 in the previous processing in the same hierarchy (n) where this is the target block PBs. It is calculated and stored in the buffer at the time of starting this processing.

  In the high-frequency AC component prediction of the target block PBs, as shown in Equation 7, the high-frequency AC component AChigh = {αhigh, of the target block PBs is obtained by weighted prediction using the high-frequency AC components AC (t) high and AC (l) high. βhigh, γhigh} is predicted. Then, the high-frequency AC component AChigh predicted by the high-frequency AC component prediction unit 32 is supplied to the prediction processing unit 33. Here, the constant k in Equation 7 represents the strength of prediction in weighted prediction. If this constant k is too large, it diverges, so it is necessary to set k <1. In the present embodiment, k = 1/4 is set as an example, but other appropriate values such as 1/2, 1/8, etc. may be set.

(Formula 7)
αhigh = (α (t) high−α (l) high) · k
βhigh = (β (l) high-β (t) high) · k
γhigh = 0

  In this weighted prediction, the calculated high-frequency AC components AC (t) high and AC (l) high are not simply added, but according to the prediction direction of the processed blocks PBt and PBl with the target block PB as a reference. The signs of the high-frequency AC components AC (t) high and AC (l) high are appropriately reversed.

  FIG. 12 is an explanatory diagram of sign inversion in the weighted prediction of the horizontal component (αhigh). As shown in FIG. 2, in the horizontal Hadamard base (α base), +1 and −1 are arranged in the horizontal direction. Therefore, with respect to the processed block PBt close to the target block PBs in the vertical direction (upward), it is possible to use the value (+ α (t) high) as it is without changing the sign of the AC component. And the error becomes smaller. For the processed block PBl close to the horizontal direction (left side) of the target block PBs, the pixel using the value (-α (l) high) obtained by inverting the sign of the AC component is used as the pixel on the left side. The error becomes smaller. Therefore, if the weighted prediction with sign inversion shown in Equation 7 is performed, the horizontal high-frequency AC component αhigh related to the target block PBs can be appropriately predicted.

  FIG. 13 is an explanatory diagram of sign inversion in weighted prediction of the vertical component (αhigh). As shown in FIG. 2, in the vertical Hadamard base (β base), +1 and −1 are aligned in the vertical direction. Therefore, with respect to the processed block PBt close to the target block PBs in the vertical direction (upward), the error with the upper pixel is better when the value (−β (t) high) obtained by inverting the sign of the AC component is used. Becomes smaller. For the processed block PBl close to the target block PBs in the horizontal direction (left side), the AC component is left as it is without changing the sign (+ β (l) high). The error with the other pixel becomes smaller. Therefore, if the weighted prediction with sign inversion shown in Equation 7 is performed, the high-frequency AC component βhigh in the vertical direction related to the target block PBs can be appropriately predicted.

  As another method for predicting the high-frequency AC component, only one of the vertical (upper) high-frequency AC component AC (t) high and the horizontal high-frequency AC component AC (l) high is used, and the target block PBs is used. The high-frequency alternating current AChigh may be predicted. For example, when only the left component AC (l) high is used, AC (t) high = 0 is simply set, and the coefficient k may be changed from 1/4 to 1/2, for example. Furthermore, in the present embodiment, γ component is set to γlow and γhigh = 0, but prediction of the γ component may be added using an appropriate method. Details of the prediction of the γ component are disclosed in, for example, Japanese Patent Application No. 2010-109079 already filed by the present applicant. In summary, first, an α component and a β component are calculated by linear prediction using a DC component and reference information of the target block PBs. Next, with respect to the target block PBs, a γ component is calculated by weighted addition using a DC component of a predetermined processed block and the previously calculated α and β components. The weighted addition is performed using a filter that minimizes the error of the pixel value between the target block PBs and the processed block, and the filter coefficient of the square error of the pixel value between these blocks is an extreme value or It is set in advance as a value near the extreme value.

  As shown in Formula 8, the prediction processing unit 33 adds the low-frequency AC component AClow predicted by the low-frequency AC component prediction unit 31 and the high-frequency AC component AChigh predicted by the high-frequency AC component prediction unit 32. Thus, the predicted value ACpred = {αpred, βpred, γpred} of the AC component of the target block PBs is calculated. The predicted value ACpred calculated by the prediction processing unit 33 is supplied to the AC component restoration unit 35.

(Formula 8)
αpred = αlow + αhigh
βpred = βlow + βhigh
γpred = γlow + γhigh

  The AC component restoration unit 35 includes a subtractor 35a, an irreversible conversion unit 35b, an inverse conversion unit 35c, and an adder 35d. The AC component restoration unit 35 has two roles: calculation / conversion of the residual component ΔAC generated between the AC component predicted value ACpred and the original image, and AC component restoration.

  Specifically, the subtractor 35a calculates the predicted value ACpred = {αpred. βpred, γpred} and the true value TAC = {Tα, Tβ. The difference from Tγ} is calculated as a residual component ΔAC = {Δα, Δβ, Δγ}.

(Formula 9)
Δα = Tα-αpred
Δβ = Tβ-βpred
Δγ = Tγ-γpred

  Here, the true value TAC of the AC component is calculated and supplied by the Hadamard transform unit 34. That is, among the TDC (n) calculated by the DC calculation unit 1 shown in FIG. 3, the one related to the target block PBs (four values) is read from the buffer and subjected to Hadamard transform to thereby read the target block PBs. A true value TAC of the alternating current component is generated. The closer the predicted value ACpred is to the original image (true value TAC), the smaller the residual component ΔAC. If the residual component ΔAC is reduced, the statistically observed appearance frequency of the residual component ΔAC tends to be biased near 0, which is advantageous in performing entropy coding.

  The irreversible conversion unit 35b performs irreversible conversion such as quantization on the residual component ΔAC calculated by the subtractor 35a, and outputs the result to the entropy encoding unit 36 and the inverse conversion unit 35c. The entropy coding unit 36 performs entropy coding such as Huffman coding or arithmetic coding on the data obtained by irreversibly transforming the residual component ΔAC, thereby compressing the compressed image data (compressed ACn) as the final output. ) Is generated.

  On the other hand, the inverse transform unit 35c performs a process opposite to the irreversible transform in the reversible transform unit 35b on the data obtained by irreversibly transforming the residual component ΔAC, so that the residual component ΔAC ′ = {Δα ′, Δβ ′, Δγ ′} is calculated. Although this residual component ΔAC ′ is a value obtained by restoring the original residual component ΔAC, the original residual component ΔAC is obtained because of irreversible transformation (the original value cannot be completely restored). Is slightly different.

  As shown in Equation 10, the adder 35d adds the residual component ΔAC ′ to the AC component predicted value ACpred supplied from the prediction processing unit 33, thereby restoring the AC component restoration value AC ′ = {α ′. , Β ′, γ ′}. The reason why the original residual component ΔAC is not used as the input of the adder 35d and the restored residual component ΔAC ′ is used is to prevent errors from accumulating due to repetitive processing at the time of decoding and ensure the reproducibility of the expanded image. It is to do.

(Formula 10)
α '= αpred + Δα'
β '= βpred + Δβ'
γ '= γpred + Δγ'

  The inverse Hadamard transform unit 37 receives the AC component AC ′ (restoration value) of the target block PBs output from the adder 35d and the DC component DC (n−1) of the target block PBs read from the buffer. The inverse Hadamard transform is performed, and the average pixel value DC (n) (the DC component of the sub block) of each sub block obtained by subdividing the target block PBs is calculated. The calculated direct current component DC (n) is temporarily stored in the buffer. The direct current component DC (n) stored in the buffer is read as needed in the processing in the immediately lower hierarchy and used as information necessary for the processing.

  The high-frequency AC component calculation unit 38 calculates the difference between the AC component AC ′ restored by the AC component restoration unit 35 and the low-frequency AC component AClow predicted by the low-frequency AC component prediction unit 31 as shown in Expression 11. Based on this, the high-frequency AC component AChigh (n) = {αhigh (n), βhigh (n), γhigh (n)} of the target block PBs is calculated.

(Formula 11)
αhigh (n) = α'−αlow
βhigh (n) = β'-βlow
γhigh (n) = γ'−γlow

  The high-frequency AC component AChigh (n) calculated by the high-frequency AC component calculation unit 38 is temporarily stored in the buffer. The high-frequency AC component AChigh (n) stored in the buffer is read as needed in the next and subsequent processes in the same hierarchy, and is used as processed block information necessary for the process.

  As is clear from the above description, in the current process for the target block PBs, two types of high-frequency AC components AChigh and AChigh (n) are calculated, but both are used for different purposes. Please keep in mind. The high-frequency AC component AChigh that is the output of the high-frequency AC component prediction unit 32 is a value that should be reflected in the predicted value ACpred itself of the target block PBs calculated in the current process. On the other hand, the high-frequency AC component AChigh (n), which is the output of the high-frequency AC component calculation unit 38, predicts the high-frequency AC component AChigh in the subsequent processing, in other words, the prediction residual of the target block PBs. Used to reflect ΔAC (≈ΔAC ′) in other blocks.

  As described above, according to the RACP encoder according to the present embodiment, the predicted value ACpred of the AC component related to the target block PBs is not the low-frequency AC component AClow itself predicted by the low-frequency AC component prediction unit 31 but the high-frequency component AClow. The corrected value is used by adding the high frequency AC component AChigh predicted by the AC component prediction unit 32. That is, the predicted value ACpred is calculated by correcting the low-frequency AC component AClow with the high-frequency AC component AChigh. The high-frequency AC component prediction unit 32 has already calculated the high-frequency AC component AChigh (which cannot be predicted by the low-frequency AC component prediction) to be taken into account in the current prediction by the high-frequency AC component AChigh (n) of the processed block, that is, the previous processing. It is calculated by weighted prediction using the determined value. The high-frequency AC component AChigh (n) corresponds to the restoration value AC (≈AC ′) obtained by adding the residual component ΔAC (≈ΔAC ′) to the predicted value ACpred and the low-frequency AC component AClow is removed. Therefore, the high-frequency AC component AChigh (n) depends on the residual component ΔAC, and the higher the residual component ΔAC with the original image becomes, the higher the high-frequency AC component AChigh (n) becomes. The current high-frequency AC component AChigh is predicted from the previous high-frequency AC component AChigh (n), and this is added to the current low-frequency AC component AClow. As a result, the previous residual component ΔAC according to the current predicted value ACpred is reflected. Also, the next high frequency AC component AChigh is predicted using the high frequency AC component AChigh (n) calculated by the high frequency AC component calculation unit 38 in this processing, and this is added to the next low frequency AC component AClow. . As a result, the current residual component ΔAC ′ (past value) is reflected in the predicted value ACpred (future value) after the next time. By repeating such processing, the residual component ΔAC of a certain block is sequentially propagated and reflected to other blocks as a high-frequency AC component AChigh (n).

  The propagation of the residual component ΔAC will be described in detail with reference to FIG. FIG. 14 is an explanatory diagram of residual propagation starting from an arbitrary block PB11. The residual component ΔAC11 of the block PB11 is reflected in the predicted value ACpred of the block PB12 via the high-frequency AC component AChigh of the block PB12 that refers to the block B11, and then also to the subsequent horizontal blocks PB13 and PB14 Propagated sequentially. On the other hand, the residual component ΔAC11 is reflected in the predicted value ACpred of the block PB21 via the high-frequency AC component AChigh of the block PB21 referring to the block PB11, and then sequentially to the subsequent vertical blocks PB31 and PB41. Propagated. Similarly, the residual component ΔAC11 is sequentially propagated to the horizontal block located on the right side of each of the blocks PB21, PB31, and PB41. Here, due to the use of weighted prediction as a prediction method for the high-frequency AC component AChigh, the degree of influence of the residual component ΔAC11 on other blocks decreases as the distance from the block PB11 increases. This is because, from Equation 7 described above, as the block affected by the residual component ΔAC becomes farther from the reference block PB11, the influence of other reference blocks closer to this block becomes stronger. The fact that an image has a spatial correlation means that a spatial correlation is also recognized in the way in which prediction errors occur. Based on this knowledge, the closer to the block PB11, the stronger the residual component ΔAC11 is reflected in the predicted value ACpred. As a result, the residual component ΔAC that would similarly occur in an image area that is spatially close by only normal AC component prediction (low-frequency AC component prediction) can be canceled by adding the high-frequency AC component AChigh. How strongly the residual component ΔAC11 is reflected in other blocks is determined by a coefficient k (see Equation 7) representing the strength of prediction. On the contrary, as the block is farther from the block PB11, the degree of reflection of the residual component ΔAC11 with respect to the predicted value ACpred is reduced. Thereby, the adverse effect of the residual component ΔAC11 on the block having low correlation is suppressed. Residual propagation as described above occurs in the same manner for blocks other than the block PB11.

  Due to such propagation of the residual component ΔAC, a component that is not captured only by the AC component prediction (the one resulting from the previous residual component ΔAC) is added to the predicted value ACpred, so that the predicted image tends to approach the original image As a result, the prediction accuracy can be further improved. Then, the image compression rate can be increased by the amount that the prediction accuracy is improved. In particular, if pipeline parallel processing as shown in FIG. 4 is performed, the overall processing time can be shortened.

(RACP encoding program)
Next, a description will be given of a RACP encoding program for realizing processing equivalent to a RACP encoder realized as hardware by software processing of a computer. Since there is no essential difference between hardware processing and software processing, only a brief description is given here, and the details thereof refer to the above-described description of the RACP encoder.

  FIG. 15 is a flowchart of the RACP encoding program. In software processing by a computer, sequential processing as shown in FIG. 5 is fundamental. First, in step 1, a plurality of blocks (4 types) are set on the image plane of the input image, and average pixel values TDC0 to TDC3 of the respective blocks are calculated. The calculated average pixel values TDC0 to TDC3 are stored in the buffer as needed. When all the blocks have been processed, the process proceeds to Step 2.

  In Step 2, DC coding, that is, differential pulse width modulation and entropy coding is performed on the average pixel value TDC0 of the 8 × 8 block, and thereby the DC component of the image that becomes part of the compressed data. DC0 is generated and output. The restored value is stored in the buffer as DC0.

  In step 3, the hierarchy number LN is set to 1 (initial value). LN = 1 indicates that the hierarchy to be processed is the highest hierarchy. Each time one hierarchical process is completed, it is incremented by 1 (step 16), and the process proceeds in order to the lower hierarchy. Then, when LN = 3, that is, when the processing of the lowest layer is finished, all the processing is finished (step 15).

  In step 4, the block number BN is set to 1 (initial value). With the setting of the layer number LN in the previous step 3, the size of the block to be processed in that layer is uniquely specified, and the end block number BNend corresponding to the total number of blocks is also specified. Each time processing of a block in the same hierarchy is completed, the block number BN is incremented by 1 (step 14), and the processing target is sequentially shifted in a predetermined direction. Then, when the processing of the block corresponding to the end block number BNend is completed, the processing in the hierarchy is completed (step 13).

  In step 5, the low-frequency AC component AClow is predicted with the block designated by the block number BN as the target block PBs. The low-frequency AC component AClow is calculated by AC component prediction using the DC component DC (n−1) (calculated at the immediately higher hierarchy) of the block corresponding to the reference region of the target block PBs. Note that not only AC component prediction using only blocks of the same size as the target block PBs, but also AC component prediction using the DC component DC (n) (calculated in the same hierarchy) of the sub-block together may be performed. As described above.

  In step 6, the high-frequency AC component AChigh of the target block PBs is predicted. The high-frequency AC component AChigh is calculated by weighted prediction using the high-frequency AC component AChigh (n) (calculated in the same hierarchy) of the processed block that is located in the vicinity of the target block PBs and has already been processed in the previous processing. Is done. By this prediction, the residual component ΔAC ′ of the processed block is reflected in the high-frequency AC component AChigh.

  In step 7, the predicted value ACpred of the AC component of the target block PBs is calculated by adding the low frequency AC component AClow of the target block PBs and the high frequency AC component AChigh of the target block PBs. As a result, the residual component ΔAC ′ of the processed block described in step 6 is reflected in the predicted value ACpred via the high-frequency AC component AChigh, so that improvement in prediction accuracy can be expected. In step 8, the difference between the predicted value ACpred of the target block PBs and the true value TAC of the original image corresponding to the target block PBs is calculated as a residual component ΔAC.

  In Step 9, the lossy component ΔAC calculated in Step 8 is subjected to irreversible transformation and entropy coding. As a result, compressed AC (n) is generated as part of the compressed image data. In step 10 following step 9, the residual component ΔAC subjected to the irreversible transformation is inversely processed to calculate a restored value ΔAC ′ obtained by restoring the original residual component ΔAC.

  In step 11, the AC component AC ′ of the target block PBs is restored by adding the residual component ΔAC ′ (restoration value) to the predicted value ACpred calculated in step 7. Then, the average pixel value DC (n) of each sub-block obtained by subdividing the target block PBs is calculated by inverse Hadamard transformation using this as an input.

  In step 12, based on the difference between the AC component restoration value AC ′ calculated in step 11 and the low-frequency AC component AClow calculated in step 5, it is used as processed block information in the subsequent processing. A high-frequency AC component AChigh (n) is calculated. Note that the processing in steps 10 to 12 is skipped in LN = 3, that is, in the lowest hierarchical processing.

  In step 13, it is determined whether or not the block number BN has reached the end block number BNend. Until the end block number BNend is reached, the block number BN is incremented by 1 in step 14, and the loop of steps 5 to 14 is repeated. When the end block number BNend is reached, that is, when the processing of all the blocks in the image plane is completed, the process exits the loop and proceeds to step 15.

  In step 15, it is determined whether or not the hierarchy number LN has reached 3 (the lowest hierarchy). Until reaching LN = 3, the layer number LN is incremented by 1 in step 16, and the loop of steps 5 to 14 is repeated as the processing of each layer. As a result, the size of the block to be processed gradually decreases as the hierarchy becomes lower, and the series of processes described above are recursively executed. The DC component DCn (output) of the sub-block calculated on the upper layer side is used as the DC component DCn-1 (input) of the block on the lower layer side. When LN = 3 and BN = BNend is reached, that is, when all the processes in the lowest hierarchy are completed, the loop exits from the determination of steps 13 and 15, whereby the processing of this program Ends.

  According to the RACP encoding program according to the present embodiment, the prediction accuracy can be further improved and the compression rate of the image can be increased, as in the above-described RACP encoder.

(RACP decoder)
FIG. 16 is an overall configuration diagram of a RACP decoder that decompresses compressed data generated by the above-described RACP encoder or RACP encoding program. This decoder is mainly composed of DC decoding 5 and three hierarchical processing units 6a to 6c. Data DCn (n = 0, 1, 2, 3) output from these units 5, 6a to 6c is temporarily stored in a buffer (storage unit) (not shown).

  The DC decoding 5 and the hierarchy processing units 6a to 6c are configured to have 8 × 8 blocks (the highest layer), 4 × 4 blocks (the second layer), 2 × 2 blocks (the third layer), and 1 × 1 block ( In the hierarchical structure of the lowest hierarchy, the hierarchical processing assigned to itself is performed. The DC decoding unit 5 in the highest hierarchy generates an average pixel value DC0 of 8 × 8 blocks by performing reverse processing of the processing applied at the time of image compression on the compressed data related to the DC component DC0 of the image. This is supplied to the hierarchy processing unit 6a of the second hierarchy.

  The hierarchical structure itself of the RACP decoder is substantially the same as that of the RACP encoder, but the AC components AC1 to AC3 that are the outputs of the hierarchical processing units 3a to 3c in the encoder are input to the hierarchical processing units 6a to 6c in the decoder. Is different. These hierarchical processing units 6a to 6c restore the average pixel value DCn based on the compressed data related to the AC components AC1 to AC3 of the image and the average pixel value DCn-1 supplied from the immediately higher hierarchy. The restored average pixel value DCn is output for supply to the immediately lower hierarchy as necessary, and is stored in the buffer. Then, a set of average pixel values DC3 of 1 × 1 blocks calculated by the lowest-level hierarchy processing unit 6c is the final decompressed image. Note that the direction of sequential shift in the decoder conforms to that of the encoder. The decoder processing may be either parallel processing or sequential processing.

  FIG. 17 is a configuration diagram of the hierarchy processing unit 6 (generic name of 6a to 6c) in the RACP decoder. The hierarchical processing units 6 in the respective layers have substantially the same basic configuration and operation except that the sizes of the blocks to be handled are different (the lowermost hierarchical processing unit 6c can be partially simplified). The hierarchical processing unit 6 includes a low-frequency AC component prediction unit 61, a high-frequency AC component prediction unit 62, a prediction processing unit 63, an AC component restoration unit 65, an inverse Hadamard transform unit 67, and a high-frequency AC component calculation unit 68. It is mainly composed.

  The low-frequency AC component prediction unit 61 predicts the low-frequency AC component of the target block PBs by AC component prediction using the DC component DC (n−1) of the block corresponding to the reference region of the target block PBs. The high-frequency AC component prediction unit 62 is positioned in the vicinity of the target block PBs and is subjected to weighted prediction using the high-frequency AC component AChigh (n) of the processed block that has already been processed in the previous processing, so that the high-frequency AC component prediction unit 62 Predict AC component AChigh. The prediction processing unit 63 adds the low-frequency AC component AClow predicted by the low-frequency AC component prediction unit 61 and the high-frequency AC component AChigh predicted by the high-frequency AC component prediction unit 62 to thereby add the target block PBs. The predicted value ACpred of the AC component is calculated. The processing so far is the same as the processing of the units 31 to 33 of the RACP encoder shown in FIG.

  The AC component restoring unit 65 adds the predicted value ACpred calculated by the prediction processing unit 63 and the residual component ΔAC ′ (restored value) of the original image, thereby restoring the AC component restored value AC of the target block PBs. 'Is calculated. The AC component restoration unit 65 includes an inverse conversion unit 65c similar to the inverse conversion unit 35c illustrated in FIG. 6 and an adder 65d similar to the adder 35d illustrated in FIG. There is no such thing as the device 35a or the irreversible conversion unit 35b. The inverse transform unit 65c performs inverse processing on both the irreversible transform and the entropy encoding performed at the time of image compression on the compressed data (compression AC (n)) of the irreversibly compressed image, thereby performing the original processing. The residual component ΔAC ′ with the image is restored. The adder 65d calculates the restored value AC ′ of the AC component of the target block PBs by adding the residual component ΔAC ′ output from the inverse transform unit 65c to the predicted value ACpred calculated by the prediction processing unit 63. To do.

  The inverse Hadamard transform unit 67 receives the AC component AC ′ (restoration value) of the target block PBs output from the adder 65d and the DC component DC (n−1) of the target block PBs read from the buffer. The inverse Hadamard transform is performed, and the average pixel value DC (n) (the DC component of the sub block) of each sub block obtained by subdividing the target block PBs is calculated. The calculated direct current component DC (n) is temporarily stored in the buffer. The direct current component DC (n) stored in the buffer is read as needed in the processing in the immediately lower hierarchy and used as information necessary for the processing.

  The high frequency AC component calculation unit 68 determines the target block PBs based on the difference between the AC component AC ′ restored by the AC component restoration unit 65 and the low frequency AC component AClow predicted by the low frequency AC component prediction unit 31. The high frequency alternating current component AChigh (n) is calculated. The calculated high-frequency AC component AChigh (n) is temporarily stored in the buffer. The high-frequency AC component AChigh (n) stored in the buffer is read as needed in the next and subsequent processes in the same hierarchy, and is used as processed block information necessary for the process.

  As described above, according to the RACP decoder according to the present embodiment, the compressed data generated by the above-described RACP decoder or RACP encoding program can be appropriately decompressed. In particular, if pipeline parallel processing as shown in FIG. 4 is performed, the overall processing time can be shortened.

(RACP decoding program)
Next, a RACP decoding program for realizing processing equivalent to the RACP decoder realized as hardware by software processing of a computer will be described. Since there is no essential difference between hardware processing and software processing, only a brief description will be given here, and details thereof will be referred to the above-described description of the RACP decoder and the like.

  FIG. 18 is a flowchart of the RACP decoding program. In software processing by a computer, sequential processing as shown in FIG. 5 is fundamental. First, in step 21, the DC data DC0 compressed data of the image is subjected to DC decoding, that is, the reverse processing of the encoding processing performed at the time of data compression, thereby the average pixel of the 8 × 8 block. The value DC0 is restored and stored in the buffer.

  In steps 22 and 23, similarly to the RACP encoding program, the layer number LN is set to 1 (initial value) and the block number BN is set to 1 (initial value). Next, in steps 24 and 25, the block specified by the block number BN is set as the target block PBs, and the low-frequency AC component AChigh and the high-frequency AC component AChigh are predicted. In step 26 following step 25, the predicted value ACpred of the AC component of the target block PBs is calculated by adding the low frequency AC component AClow of the target block PBs and the high frequency AC component AChigh of the target block PBs.

  In step 27, the compressed ACn (n) that is the compressed data of the image is decompressed, and thereby the residual component ΔAC ′ is restored. In step 28, the AC component restoration value AC 'of the target block PBs is restored by adding the residual component ΔAC' calculated in step 27 to the AC component prediction value ACpred calculated in step 26. . Then, the average pixel value DC (n) of each sub-block obtained by subdividing the target block PBs is calculated by inverse Hadamard transformation using this as an input. Then, in step 29, based on the difference between the AC component restoration value AC ′ calculated in step 18 and the low-frequency AC component AClow calculated in step 24, the processed block information is processed in the subsequent processing. A high-frequency AC component AChigh (n) to be used is calculated. Note that the processing in step 29 is skipped in LN = 3, that is, in the lowest hierarchical processing.

  In step 30, it is determined whether or not the block number BN has reached the end block number BNend. Until the end block number BNend is reached, the block number BN is incremented by 1 in step 31, and the loop of steps 24-31 is repeated. When the end block number BNend is reached, that is, when the processing of all the blocks in the image plane is completed, the process exits the loop from the determination result of step 30, and proceeds to step 32.

  In step 32, it is determined whether or not the hierarchy number LN has reached 3 (the lowest hierarchy). Until N = 3, the layer number LN is incremented by 1 in step 33, and the loop of steps 24-31 is repeated as the processing of each layer. As a result, the size of the block to be processed gradually decreases as the hierarchy becomes lower, and the series of processes described above are recursively executed. The DC component DCn (output) of the sub-block calculated on the upper layer side is used as the DC component DCn-1 (input) of the block on the lower layer side. When LN = 3 and BN = BNend is reached, that is, when the processing of the lowest layer is completed, the loop exits from the determination of steps 30 and 32, whereby the processing of this program is executed. finish.

  According to the RACP decoding program according to the present embodiment, the same effects as those of the above-described RACP decoder can be obtained.

  In the above description of the present embodiment, the high-frequency AC component AChigh caused by the previous residual component ΔAC cannot be predicted only by the AC component prediction, and is a component that is not included in the predicted value AClow. Therefore, it is positioned as a high frequency component for the predicted value AClow (low frequency component). If this is grasped from another viewpoint, the high-frequency AC component AChigh can be regarded as a correction component of the AC component AClow predicted by the AC component prediction. In this case, the process of calculating the predicted value ACpred by adding the low AC frequency component AClow and the high frequency AC component AChigh is performed by converting the predicted value AClow predicted by the AC component prediction into the previous residual component ΔAC (≈AChigh (n) In other words, it is a process for calculating the corrected predicted value ACpred.

  As described above, the AC component prediction method according to the present invention can be widely applied to applications that increase the compression rate of an image by improving prediction accuracy.

DESCRIPTION OF SYMBOLS 1 DC calculation part 2 DC encoding part 3 (3a-3c) Hierarchical process part 5 DC decoding part 6 (6a-6c) Hierarchical process part 31, 61 Low frequency alternating current component prediction part 32, 62 High frequency alternating current component prediction part 33 , 63 Prediction processing unit 34 Hadamard transform unit 35, 65 AC component restoration unit 35a Subtractor 35b Non-reversible transform unit 35c, 65c Inverse transform unit 35d, 65d Adder 36 Entropy encoding unit 37, 67 Inverse Hadamard transform unit 38, 68 High-frequency AC component calculator

Claims (12)

In the AC component prediction system that performs AC component prediction in units of blocks by iterative processing for each block set on the image plane,
A low-frequency AC component prediction unit that predicts a low-frequency AC component of the target block by AC component prediction using a DC component of a block corresponding to the reference region of the target block to be processed in the current process;
A high-frequency AC component prediction unit that predicts a high-frequency AC component of the target block by weighted prediction using a high-frequency AC component of a processed block that is located in the vicinity of the target block and has already been processed in a previous process;
Prediction for calculating the predicted value of the AC component of the target block by adding the low-frequency AC component predicted by the low-frequency AC component prediction unit and the high-frequency AC component predicted by the high-frequency AC component prediction unit A processing unit;
An AC component restoration unit that calculates a restoration value of the AC component of the target block by adding a prediction value calculated by the prediction processing unit and a residual component of the original image;
Based on the difference between the restoration value calculated by the AC component restoration unit and the low frequency AC component predicted by the low frequency AC component prediction unit, the high frequency used as information of the processed block in the subsequent processing An AC component prediction system comprising: a high-frequency AC component calculation unit that calculates an AC component.   The high-frequency AC component prediction unit performs the processing according to a prediction direction of the processed block with reference to the target block so that an error of a pixel value between the target block and the processed block becomes small. 2. The AC component prediction system according to claim 1, wherein the weighted prediction is performed after inverting the sign of a high-frequency AC component of a completed block. The alternating current component includes a horizontal component that defines a horizontal Hadamard basis component,
The high-frequency AC component prediction unit inverts the sign of the horizontal component related to the processed block close to the target block in the vertical direction and the horizontal component related to the processed block close to the target block in the horizontal direction. The AC component prediction system according to claim 2, wherein the weighted prediction regarding the high-frequency AC component included in the horizontal component is performed by adding a value. The alternating current component includes a vertical component that defines a vertical Hadamard basis component,
The high-frequency AC component prediction unit is configured to invert the sign of the vertical component related to the processed block close to the target block in the vertical direction and the vertical related to the processed block close to the target block in the horizontal direction. 4. The AC component prediction system according to claim 2, wherein the weighted prediction relating to the high-frequency AC component included in the vertical component is performed by adding the component. The AC component prediction system is an encoder for irreversibly compressing an image,
The AC component restoration unit is
A subtractor that calculates a difference between a prediction value calculated by the prediction processing unit and a true value in the original image as a residual component;
An irreversible conversion unit that performs irreversible conversion on the residual component calculated by the subtractor;
An entropy encoding unit that performs entropy encoding on the residual component output from the lossy conversion unit and outputs compressed image data;
An inverse transform unit that restores the residual component by performing a process reverse to the irreversible transform on the residual component output from the irreversible transform unit;
5. The AC component according to claim 1, further comprising an adder that adds the residual component output from the inverse transform unit to the prediction value calculated by the prediction processing unit. Prediction system. The AC component prediction system is a decoder that decompresses a lossy compressed image,
The AC component restoration unit is
An inverse transform unit that restores a residual component from the original image by performing inverse processing of lossy transformation and entropy encoding performed at the time of image compression on the compressed data of the lossy compressed image;
5. An adder that restores an AC component by adding a residual component output from the inverse transform unit to a predicted value calculated by the prediction processing unit. An AC component prediction program described in any one of the above. In the AC component prediction program that causes the computer to execute AC component prediction performed in units of blocks by iterative processing for each block set on the image plane,
A first step of predicting a low-frequency AC component of the target block by AC component prediction using a DC component of a block corresponding to a reference region of the target block to be processed in the current process;
A second step of predicting the high-frequency alternating current component of the target block by weighted prediction using the high-frequency alternating current component of the processed block located in the vicinity of the target block and already processed in the previous processing;
A third step of calculating a predicted value of the AC component of the target block by adding the low-frequency AC component and the high-frequency AC component;
A fourth step of calculating a restoration value of the AC component of the target block by adding the prediction value and a residual component of the original image;
Based on the difference between the restoration value and the low-frequency AC component, a fifth step of calculating a high-frequency AC component used as information of the processed block in the subsequent processing is executed. AC component prediction program.   In the second step, the processed block is processed according to a prediction direction of the processed block with respect to the target block so that an error of a pixel value between the target block and the processed block becomes small. The AC component prediction program according to claim 7, wherein the weighted prediction is performed after inverting the sign of the high-frequency AC component of the block. The alternating current component includes a horizontal component that defines a horizontal Hadamard basis component,
The second step is a value obtained by inverting the sign of the horizontal component related to the processed block close to the target block in the vertical direction and the horizontal component related to the processed block close to the target block in the horizontal direction. The AC component prediction program according to claim 8, wherein the weighting prediction is performed on the high-frequency AC component included in the horizontal component by adding. The alternating current component includes a vertical component that defines a vertical Hadamard basis component,
The second step includes a value obtained by inverting the sign of the vertical component related to the processed block adjacent to the target block in the vertical direction and the vertical component related to the processed block adjacent to the target block in the horizontal direction. 10. The AC component prediction program according to claim 8, wherein the weighted prediction is performed on the high-frequency AC component included in the vertical component by adding. The AC component prediction program is an encoding program for irreversibly compressing an image,
The fourth step includes
Calculating a difference between the predicted value and a true value in the original image as a residual component;
Applying an irreversible transformation to the residual component;
Performing entropy coding on the residual component and outputting compressed image data;
Restoring the residual component by subjecting the residual component to a process opposite to the lossy transformation;
The AC component prediction program according to claim 7, further comprising a step of adding the residual component to the prediction value. The AC component prediction program is a decoding program for expanding an irreversible compressed image,
The fourth step includes
Restoring the residual component from the original image by performing reverse processing of the lossy transformation and entropy encoding performed at the time of image compression on the compressed data of the lossy compressed image;
The AC component prediction program according to claim 7, further comprising a step of restoring an AC component by adding the residual component to the predicted value.

Images