JP5652101B2 - Image processing apparatus and image processing method - Google Patents

Description

  The present disclosure relates generally to an image processing apparatus and method, and more particularly to an image processing apparatus and method that performs encoding or decoding.

  Image lossless encoding methods represented by PNG (Portable Network Graphics) and JPEG (Joint Photographic Experts Group) -LS do not deteriorate the image quality during compression, and mainly compress images that require high definition. Has been adopted. In these still image reversible compressions, the pixel value of the compression target pixel is predicted from the surrounding pixels, and the redundancy of the image is effectively utilized by taking the difference between the predicted value and the actual pixel value. In order to compress the residual generated by prediction at a high compression rate, Huffman coding using a variable length code table constructed based on the probability distribution of the generated prediction residual is effective. In still image compression using the Huffman code, a prediction value is generated before compression, a residual with a compression target pixel is calculated, and a distribution of the residual is calculated. A Huffman code table is generated by constructing a Huffman tree indicating the appearance probability of each difference value based on the distribution. Thereafter, compressed data is generated by the generated Huffman code table. In this method, when the compressed data is expanded, the code table used for compression must be known. Therefore, the compressed data is expanded by adding not only the compressed residual data but also the code table itself. Is possible. Since the code table itself is added to the compressed data, the size of the compressed data increases, but in the case of an image size of a certain level or more, the size of the code table in the compressed data does not actually matter. For example, when compressing the brightness of an image of 1920 × 1088 pixels, if one pixel is 8 bits and the compression rate is 50%, the size of the compressed data is about 1 Mbyte. On the other hand, even when the maximum length of the variable length code is 16 bits, the size required to have a code for all the differences is at most about 1 Kbyte, which is about 0.1% of the size of the compressed data. It can be said that there is virtually no loss.

  Thus, although the code table size is sufficiently small with respect to the compressed data, the entire image must be scanned once in order to construct the Huffman code table. For this reason, a memory that can fit the entire screen is required as the input image storage memory. Also, since encoding cannot be started unless the entire image is scanned once, sequential encoding is difficult, and it is difficult to increase the speed in terms of encoding speed. In order to realize the sequential encoding, a method of using a code table prepared in advance without scanning one entire screen can be considered. However, since the code table in this case is not a Huffman code corresponding to the probability distribution of the information source, the compression rate is inferior to Huffman coding that achieves optimum efficiency. For this reason, it is a common practice to divide an image into predetermined blocks and select a block with the highest compression rate from a plurality of code tables. In this case, since each block is compressed independently, the memory required for code table selection and encoding may be a size that can accommodate a predetermined unit of block, and can be sequentially encoded in units of blocks. . In this configuration, a code table prepared in advance is based on a specific general-purpose probability density function such as a Laplace distribution or a normal distribution. This is based on the fact that the distribution of predicted residuals is generally 0, and the distribution is symmetrical with the appearance frequency decreasing as the residual increases. As described above, when the code table is based on a general-purpose probability density function, the code table used in each block can be determined from the variation degree of the prediction residual generated in each block, that is, the variance.

  In this configuration, in order to achieve a compression rate close to that of Huffman coding, code table selection may be performed in finer block units. However, as the block size decreases, the size of the flag bit indicating the used code table increases relative to the compressed data, so the compression rate decreases. In addition, high-definition image blocks have low correlation with surrounding pixels, and the generated prediction value is often significantly different from the actual pixel value. Therefore, a code table based on a specific probability density function has a high compression ratio. Cannot be achieved. In particular, as the block size is smaller, the residual distribution is different from an expected distribution that is bilaterally symmetric at the center 0, which causes a problem that the compression ratio decreases.

JP-A-7-107492 JP 2000-115782 A

  In view of the above, it is desired to provide an image processing apparatus and method that can achieve a high compression rate even when the image area is small in the still image lossless compression method that selects a code table for a local image area.

The image processing apparatus calculates a first variance value according to the peripheral pixel value around the target pixel, and calculates a second variance value according to the predicted value of the target pixel calculated based on the peripheral pixel value. When a probability distribution in which the horizontal axis is a prediction error that is a difference between the value of the target pixel and the predicted value and a vertical axis is the appearance probability of the prediction error is considered, code table is selected according to a code table corresponding to the asymmetric probability distribution having a second variance value on the other hand have a first variance value, the first dispersion value and the second dispersion value a selecting unit, characterized in that it comprises a said encoding to encode or decode the code table said it selected residual obtained by subtracting the prediction value from the value of the target pixel or decoding unit.

In the image processing method, the image processing apparatus calculates the first variance value according to the peripheral pixel value around the target pixel , and the first processing unit according to the predicted value of the target pixel calculated based on the peripheral pixel value . calculating a second dispersion value, consider an image processing apparatus, a prediction error horizontal axis is the difference between the predicted value and the value of the target pixel, and the vertical axis represents the probability distribution is a probability of occurrence of the prediction error If the the code table corresponding to the asymmetric probability distribution having a second variance value on the other hand has a first distribution value, while the left and right, the first dispersion value and the second selected according to the dispersion value, the image processing apparatus, and characterized in that a residual obtained by subtracting the prediction value from the value of the target pixel includes the stages of encoding or decoding by the selected code table To do.

  According to at least one embodiment of the present disclosure, efficient coding can be realized and a high compression rate can be achieved by using a code table corresponding to a left-right asymmetric probability distribution having different left and right variance values.

It is a figure which shows an example of the probability distribution of the prediction error which has two peaks. It is a figure for demonstrating selection of a code table. It is a figure which shows an example of a left-right asymmetric probability distribution. It is an example of the code table corresponding to the left-right asymmetric probability distribution comprised by combining two code tables. It is a figure which shows an example of a structure of an encoding apparatus. 6 is a flowchart showing a flow of operations of the encoding device shown in FIG. 5. It is a flowchart which shows an example of the process which produces | generates an intrinsic dispersion value table. It is a figure which shows an example of the data structure of a some code table. It is a flowchart which shows the process which produces | generates a some code table. It is a figure which shows a raster scan order. It is a figure for demonstrating an example of the process which calculates | requires a predicted value and an adjacent pixel value with the pixel scanned previously in the raster scan order. It is a flowchart which shows the flow of the encoding process by the encoding apparatus shown in FIG. It is a figure which shows an example of a structure of a decoding apparatus. It is a flowchart which shows the flow of the encoding process by the decoding apparatus shown in FIG. It is a figure which shows an example of a structure of the image processing system which performs encoding and decoding.

  In reversible compression of an image, a prediction value is calculated based on pixel distribution of surrounding pixels and information when compressed in the past, and a difference between the pixel value of interest (encoding target pixel value) and the prediction value is encoded. There is a first method and a second method that simply encodes a difference between a pixel value of interest and an adjacent point. For lossless compression of natural images, the first method is said to have a smaller difference value than the second method. However, it is not necessary to generate a predicted value with a complicated calculation formula by the first method, and the difference value is sufficiently small even if the adjacent pixel value is simply used as the predicted value by the second method regardless of the surrounding state. There is a case.

  Here, the predicted value of the first method is e, and the pixel value (adjacent pixel value) of the adjacent representative point is c. The target pixel value has the highest probability of matching with the predicted value e. However, when considering the pixel distribution of a local image region (image block) around the pixel of interest, the probability that the pixel value of interest matches the predicted value e is so high in an image block having a complicated pixel distribution. Don't be. On the other hand, regardless of the past pixel value series, the probability that the adjacent pixel value c matches the target pixel value cannot be ignored. As a result, the probability distribution of the difference (prediction error) between the target pixel value and the predicted value e is 0 (position where the target pixel value matches the predicted value e) and c-e (the target pixel value is adjacent). The shape has a peak at two points (when it matches the pixel value c).

  FIG. 1 is a diagram illustrating an example of a probability distribution of prediction errors having two peaks. 1A and 1B, the horizontal axis indicates a prediction error, that is, a value obtained by subtracting the prediction value e from the pixel value of interest, and the vertical axis indicates the appearance probability of each prediction error. As shown in FIG. 1A, while there is a symmetrical probability distribution 10 centered on the position where the prediction error is zero (the position where the pixel value of interest is equal to the prediction value e), the position of the prediction error ce There is a probability distribution 11 centered on (position where the target pixel value is equal to the adjacent pixel value c). As a result, as shown in FIG. 1B, the probability distribution of the prediction error is a left-right asymmetric probability distribution 12 obtained by superimposing the probability distribution 10 and the probability distribution 11. The probability distribution 12 has a shape with peaks at two points, a 0 position and a c-e position.

  In an image block having a simple pixel distribution, if the predicted value and the adjacent pixel value are different, the probability that the adjacent pixel value matches the target pixel value is sufficiently higher than the probability that the predicted value is correct. The prediction error distribution is small and symmetrical. However, as described above, in an image block having a complicated pixel distribution, when the predicted value and the adjacent pixel value are different, the pixel value of interest matches the adjacent pixel value compared to the probability that the predicted value is correct. Probability cannot be ignored. As a result, the prediction error distribution is a bilaterally asymmetric distribution having different variance values in the positive region and the negative region. Specifically, when the neighboring pixel value is larger than the predicted value, the variance value in the positive region is larger than the variance value in the negative region, and when the neighboring pixel value is smaller than the predicted value, the negative value is negative. A probability distribution in which the variance value in the region is larger than the variance value in the positive region is obtained.

  FIG. 2 is a diagram for explaining selection of a code table. In FIG. 2, the pixel array 21 is an extracted image region including peripheral pixels a, b, and c of the pixel of interest x in the entire screen. Among these peripheral pixels, for example, the pixel c is used as an adjacent pixel. The code table set 22 includes a plurality of code tables A to E corresponding to probability distributions of different variance values. As described above, in an image block having a complicated pixel distribution, the prediction error distribution is a bilaterally asymmetric distribution having different variance values in a positive region and a negative region. In order to deal with such a complex distribution, the first variance value and the second variance value are calculated from the peripheral pixel values a to c, and the left and right asymmetrical values having different variance values on the left and right are calculated from the code table set 22. A code table corresponding to the probability distribution is selected according to the first variance value and the second variance value. A method for calculating the first variance value and the second variance value will be described in detail later.

  Each of the code tables A to E of the code table set 22 may be a code table corresponding to a probability distribution that is asymmetrical. In this case, for example, a code table corresponding to a probability distribution in which the left side of the distribution is the first variance value and the right side of the distribution is the second variance value may be selected. Alternatively, each of the code tables A to E of the code table set 22 is a code table corresponding to one of the left and right distributions of the probability distribution (for example, a code for a non-negative integer prediction error corresponding to the right one distribution). Table). These code tables A to E are a plurality of code tables corresponding to a plurality of probability distributions having a plurality of different variance values. In this case, for example, the code table B corresponding to the first variance value is selected from the code table set 22, the code table E corresponding to the second variance value is selected, and the code table B and the code table E are selected. By combining them, a code table corresponding to a left-right asymmetric probability distribution may be configured. In the example of FIG. 2, the contents of the selected code table B are shown in the table 23, and the contents of the selected code table E are shown in the table 24. Each table includes a plurality of difference values and a plurality of codes that are associated with each other on a one-to-one basis. That is, each difference value is associated with one code.

  FIG. 3 is a diagram illustrating an example of a left-right asymmetric probability distribution corresponding to a selected code table or a code table obtained by combining selected code tables. The difference (prediction error) between the target pixel value x and the predicted value e (which may be calculated from the peripheral pixel values a to c) is shown on the horizontal axis, and the appearance probability is shown on the vertical axis. Since the prediction error value c−e is in the positive region (that is, the adjacent pixel value c is larger than the prediction value e), the distribution value of the positive region distribution (the distribution value of the right distribution) is the negative region distribution. Is greater than the variance value (the variance value of the distribution on the left side).

  FIG. 4 is an example of a code table corresponding to a left-right asymmetric probability distribution configured by combining two code tables. The code table 26 corresponding to the left-right asymmetric probability distribution shown in FIG. 4 includes a table 23 corresponding to the one-sided distribution (same as the table 23 shown in FIG. 2) and a table 24 corresponding to the one-sided distribution (table 24 shown in FIG. 2). The same). However, so that the code of the table 23 and the code of the table 24 do not match, a code obtained by adding “1” to the head of the code of the table 23 shown in FIG. 2 and “0” at the head of the code of the table 24 shown in FIG. "Is used. In this example, the table 23 is associated with a negative difference value area, and the table 24 is associated with a positive difference value area. As the code of the difference value 0, a code in which “1” is added to the head of the code of the difference value 0 in the table 23 is used. As a result, the code “00” of the difference value 0 in the table 24 remains without being used. Therefore, “000” with “0” added to the head of the sign of the difference value 0 in the table 24 shown in FIG. 2 is assigned to the difference value equal to c−e. In this way, when the target pixel value x is equal to the adjacent pixel value c (that is, when the difference between the target pixel value x and the predicted value e is equal to c−e), the code table B (table 23) or the code A code corresponding to a difference value of zero in Table E (Table 24) is assigned. As a result, as shown in FIG. 1B, it is possible to assign a code having a short code length at the position of the peak that cannot be ignored, which is the cause of the distribution being asymmetrical, thereby realizing high coding efficiency. Can do.

  When the target pixel value x in FIG. 2 is encoded, a difference value x-e obtained by subtracting the predicted value e from the target pixel value x is calculated, and the difference value xe in the code table 26 in FIG. 4 corresponds to the calculated difference value. What is necessary is just to output a code | symbol. As a result, it is possible to realize efficient encoding in consideration of a case where the probability distribution of difference values (also referred to as prediction errors or residuals) is asymmetrical.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

  FIG. 5 is a diagram illustrating an example of the configuration of the encoding device. The encoding device 30 illustrated in FIG. 5 includes a code table selection unit 31, a residual generation unit 32, and an encoding unit 33. The block image data storage memory 35 stores local image block data around the target pixel to be encoded in the entire image. The encoding device 30 receives the pixel data of the image block from the block image data storage memory 35, encodes the pixel of interest, and stores the encoded data obtained by the encoding in the encoded data storage memory 36.

  Note that in FIG. 5 and the subsequent similar figures, the boundary between each functional block shown in each box and other functional blocks basically indicates a functional boundary, and separation of physical positions, It does not necessarily correspond to separation of electrical signals, separation of control logic, and the like. In the case of hardware, each functional block may be one hardware module physically separated from other blocks to some extent, or in a hardware module physically integrated with other blocks. One function may be shown. In the case of software, each functional block may be one software module logically separated from other blocks to some extent, or one function in a software module logically integrated with another block. May be shown.

  The code table selection unit 31 calculates the first variance value and the second variance value from the peripheral pixel values around the pixel of interest, and calculates the code table corresponding to the left-right asymmetric probability distribution having different variance values on the left and right. Selection is made according to the first variance value and the second variance value. The residual generation unit 32 calculates a residual by subtracting the predicted value from the target pixel value. The encoding unit 33 encodes the residual calculated by the residual generation unit 32 with the code table selected by the code table selection unit 31. The outline of the encoding process is as described with reference to FIGS. The code table selection unit 31 includes a context C1 derivation unit 41, a predicted value generation unit 42, an adjacent point selection unit 43, a context C2 derivation unit 44, a first table selection unit 45, a second table selection unit 46, and a code table construction unit. 47.

  The context C1 deriving unit 41 determines the context C1 based on the gradient of surrounding pixels. The predicted value generation unit 42 generates a predicted value of the pixel of interest based on the context C1 and surrounding pixel values. The first table selection unit 45 obtains a first variance value from the context C1, and selects a code table corresponding to the first variance value. The adjacent point selection unit 43 selects one pixel from the surrounding pixels as an adjacent pixel. The context C2 deriving unit 44 determines the context C2 based on the adjacent pixel value selected by the adjacent point selecting unit 43 and the predicted value generated by the predicted value generating unit 42. The second table selection unit 46 obtains a second variance value based on the context C2 and the first variance value supplied from the first table selection unit 45, and obtains a code table corresponding to the second variance value. select. The code table construction unit 47 combines the code table selected by the first table selection unit 45 and the code table selected by the second table selection unit 46 to construct a code table corresponding to a left-right asymmetric probability distribution.

  FIG. 6 is a flowchart showing a flow of operations of the encoding device 30 shown in FIG. First, in step S1, the context C1 deriving unit 41 calculates the context C1. In this example, as the context C1, the horizontal gradient Ghor = ab, the vertical gradient Gver = ac, and the oblique backward gradient Gnan = c + b- a is calculated. Further, the context C1 deriving unit 41 quantizes each gradient value by shifting Ghor, Gver, and Gnan to the right by 6 bits and discarding the lower 6 bits. A set of Ghor, Gver, and Gnan after the quantization is set as a context C1. Quantization is performed to reduce the total number of contexts C1.

  In step S <b> 2, the first table selection unit 45 refers to the context C <b> 1 in the inherent variance value table 50 prepared in advance to derive the intrinsic variance value of the context C <b> 1. This derivation process may be executed by the context C1 derivation unit 41 instead of the first table selection unit 45. The intrinsic dispersion value derived here basically indicates the degree of variation of surrounding pixels. A large intrinsic variance value indicates that the predicted value of the pixel value of interest is difficult to hit, that is, there is a high probability that the prediction residual will be large. However, instead of simply calculating the variance of the pixel values, the intrinsic variance value is obtained based on the gradient magnitudes in the horizontal, vertical, and diagonal directions. Therefore, for example, when the gradient in the horizontal direction is large but the gradient in the vertical direction is zero, the probability of a vertical pattern such as a vertical edge is high, and in such a case, the intrinsic dispersion value is not so large. Each value of the intrinsic dispersion value table 50 may be determined.

  FIG. 7 is a flowchart illustrating an example of processing for generating an inherent variance table. By executing the processing of FIG. 7 for various images to be evaluated, the inherent dispersion value table 50 used for the processing of FIG. 6 is generated and prepared in advance.

  In step S1, an evaluation pixel of an image to be evaluated is selected. In step S2, a prediction residual (a difference between the evaluation pixel value and the prediction value) for the evaluation pixel is derived. In step S3, a context C1 for the evaluation pixel is derived. In step S 4, the correspondence relationship between the obtained context C 1 and the prediction residual is stored in the correspondence relationship data 51. That is, the context C1 and the prediction residual are associated with each other and stored in the correspondence data 51. In step S5, it is determined whether scanning for all pixels of all images has been completed. If not completed, the process returns to step S1 and the subsequent processing is repeated.

  When scanning of all pixels is completed, one context C1 is selected in step S6. In step S7, the residual set of the selected context C1 is read out. That is, all prediction residuals associated with the selected context C1 are read out. In step S8, the variance of the read prediction residual is calculated. In step S9, it is determined whether or not a variance value has been generated for all contexts C1. If the context C1 for which no distributed value is generated remains, the process returns to step S6, and the subsequent processing is repeated.

  With the above processing, the variance of the prediction residual for each context C1 is obtained based on a number of various evaluation target images. The inherent variance value table 50 can be generated by using the variance of the prediction residual thus obtained as the inherent variance value associated with each context C1. In this way, the intrinsic dispersion value table 50 used in step S2 of FIG. 6 is prepared.

  Referring to FIG. 6 again, in step S3, the first table selection unit 45 selects a code table that matches the inherent variance value of the context C1. Specifically, a plurality of code tables corresponding to a plurality of variance values are provided, and if the first table selection unit 45 selects a code table having a variance value that is equal to or closest to the inherent variance value of the context C1. Good.

  FIG. 8 is a diagram illustrating an example of a data configuration of a plurality of code tables. A plurality of code tables are stored as a set 61 in a memory such as a nonvolatile memory. The word width is 16 bits, for example. One code table included in the code table set 61 has a data structure shown in the table 62. The table 62 includes an inherent dispersion value that is a 16-bit integer, 12-bit code data, and 4-bit code length data. The 12-bit code data includes a code having a length of 12 bits or less and “0” that fills the remaining bit positions on the right side thereof. For example, in the case of the third entry from the top, the code length is “0011”, that is, 3 bits in binary, so the actual code is “110” taking the upper 3 bits of the code data. In step S3 described above, the first table selection unit 45 selects, from the code table set 61, a code table having an eigen variance value that is equal to or closest to the eigen variance value of the context C1.

  FIG. 9 is a flowchart showing a process for generating a plurality of code tables. By executing the processing of FIG. 9, a plurality of code tables used for the processing of FIG. 6 (that is, the code table set 61 of FIG. 8) are generated and prepared in advance.

  In step S1, one intrinsic dispersion value is selected. In step S2, a probability density function having a normal distribution having the selected intrinsic variance as a variance is derived. In step S3, a Huffman code table is generated based on the generated probability density function. In step S4, it is determined whether or not a code table (the Huffman code table described above) has been generated for all eigenvariance values. If there is still an inherent variance value for which no code table has been generated, the process returns to step S1 and the subsequent processing is repeated.

  Referring to FIG. 6 again, in step S4, the predicted value generation unit 42 calculates a predicted value, the adjacent point selecting unit 43 selects an adjacent pixel, and the context C2 deriving unit 44 further calculates the predicted value and the adjacent pixel value. Based on the above, the context C2 is calculated. The predicted value generation unit 42 may select one from a plurality of predicted value generation methods according to the context C1. That is, for example, when the gradient in the horizontal direction is large but the gradient in the vertical direction is zero, the probability of a vertical pattern such as a vertical edge is high, and the value of the pixel b among the peripheral pixels a to c is set to the pixel x. It is conceivable to use a predicted value of. In the example illustrated in FIG. 6, the predicted value of the pixel of interest x is generated as a function f (a, b, c) from the peripheral pixel values a to c regardless of the content of the context C1. As an example of the function f (a, b, c), a function for obtaining an average value of three peripheral pixel values of the pixel value a, the pixel value b, and the pixel value c may be used. The prediction formula to be used is not limited to this example, and an arbitrary prediction formula may be used. The adjacent point selection unit 43 selects a pixel c adjacent to the pixel x of interest as an adjacent representative point. The adjacent pixel selection method is not limited to a method of simply selecting a specific position (for example, the horizontally adjacent pixel c), but a method of selecting an adjacent point based on edge strength as in PNG PAETH prediction, etc. May be used as appropriate. The context C2 deriving unit 44 obtains the context C2 based on the predicted value f (x, y, z) and the adjacent pixel value c. The context C2 is defined as an absolute value Diff (= | f (x, y, z) −c |) of a difference between the predicted value f (x, y, z) and the adjacent pixel value c.

  In this embodiment, the predicted value and the adjacent pixel value are obtained based on the three pixel values of the upper neighbor, the left neighbor, and the upper left neighbor, and the context C2 is calculated. It does not have to be a pixel. As long as the pixel is encoded or decoded before the target pixel (compression target pixel) in the raster scan order, the context C2 may be calculated using an arbitrary number of pixels at an arbitrary position. The same applies to the calculation of the context C1.

  FIG. 10 shows the raster scan order. In FIG. 10, an image 70 includes pixels (squares in the figure) arranged in a matrix form in the vertical and horizontal directions. Pixels arranged in a horizontal column form each pixel row 71. In the raster scan order indicated by the arrow 72, all the pixels in one pixel row 71 are scanned in order from left to right. When the right end of the image 70 is reached, the pixel 70 moves to the left end of the next pixel row 71. The pixels are scanned sequentially from left to right.

  FIG. 11 is a diagram for explaining an example of a process for obtaining a predicted value and an adjacent pixel value using pixels scanned first in the raster scan order. In FIG. 11, a 4 × 4 pixel local image region (image block) 76 including a target pixel (encoding target pixel) 75 is considered. All pixels included in the image block 76 other than the target pixel 75 are encoded in raster scan order before the target pixel 75. Therefore, if contexts C1 and C2 are obtained using the pixel values in the image block 76 and the pixel of interest 75 is encoded, the contexts C1 and C2 for decoding the pixel of interest 75 are obtained even when decoding. be able to. That is, when decoding the pixel of interest 75, the pixel values of all the pixels in the image block 76 other than the pixel of interest 75 have already been obtained in the raster scan order by the decoding process. Contexts C1 and C2 can be determined using Since the calculation of the contexts C1 and C2 for selecting the code table can be executed on the decoding process side as well as the encoding process side, the information for specifying the code table used at the time of encoding is encoded. There is no need to add to compressed data. Therefore, even if the image area unit for sequentially selecting the code table is reduced, the compression rate does not deteriorate, and the code table may be selected and switched for each pixel.

  In the example illustrated in FIG. 6, the predicted value and the adjacent pixel value are obtained based on the three pixel values of the upper neighbor, the left neighbor, and the upper left neighbor. The predicted value and the neighboring pixel value are respectively global predicted values. And can be considered as a local prediction value. That is, as shown in FIG. 11, a predicted value is obtained based on a plurality of pixels 77 and 78 (15 pixels other than the pixel of interest 75 included in the image block 76) in a relatively wide range, May be considered as a predictive value. Further, an adjacent pixel value may be obtained based on a plurality of pixels 77 in a relatively narrow range (three pixels in contact with the target pixel 75 in the image block 76), and this may be considered as a local predicted value. The adjacent pixel value in this case may be a value obtained by selecting one pixel value from the pixel values of the three pixels 77 or by calculating from the pixel values of the three pixels 77. When the global prediction value obtained in this way is at the center of the probability distribution as shown in FIG. 1 and the local prediction value (adjacent pixel value) is deviated from the center of the probability distribution, As a result, a left-right asymmetric probability distribution with a large variance on the side where the predicted value exists is obtained.

Returning to FIG. 6, in step S < ^b > 5, the second table selection unit 46 obtains the second variance value as the inherent variance value + Diff ² of the context C < ^b > 1 and selects a code table that matches the second variance value. Specifically, a plurality of code tables corresponding to a plurality of variance values are provided, and the second table selection unit 46 selects a code table having a variance value that is equal to or closest to the second variance value. That's fine.

  In step S6, the code table selected in step S3 and the code table selected in step S5 are combined, and the code table construction unit 47 constructs a code table. The method of combining the code tables is the same as described in FIG. In step S6 of FIG. 6, the code table selected in step S3 is a table Table1, and the code table selected in step S5 is a table Table2. Furthermore, a code table corresponding to a region where the difference x-e between the predicted value e (= f (a, b, c)) and the target pixel value x is positive is set as a table TableR, and the value of the difference x-e. The table corresponding to the negative region is the table TableL. When the predicted value is smaller than the adjacent pixel value, the variance becomes large in the region where the value of the difference x−e is positive (see FIG. 3), so the table Table2 having a large variance is assigned as the table TableR. On the contrary, when the predicted value> the adjacent pixel value, since the variance is large in the region where the value of the difference x−e is negative, the table Table2 having a large variance is assigned as the table TableL. When the predicted value = adjacent pixel value, the variance does not increase, and only the table Table1 is used (in this case, the table Table2 obtained in step S5 is equal to the table Table1). As described with reference to FIG. 4, a code in which “1” or “0” is added to the head of each code is output so that the code of the table Table1 does not match the code of the table Table2. .

  FIG. 12 is a flowchart showing the flow of the encoding process by the encoding device 30 shown in FIG. The process shown in FIG. 12 shows the process shown in FIG. 6 in a more organized form.

  In step S1, a context C1 is derived. In step S2, the code table Table1 is determined based on the context C1. In step S3, the predicted value calculation method f (a, b, c) is selected based on the context C1. In step S4, the context C2 is calculated based on the predicted value and the adjacent pixel value, and the code table Table2 is determined based on the context C2. In step S5, a code table corresponding to a left-right asymmetric probability distribution is constructed by combining the code table Table1 and the code table Table2. In step S6, the prediction error xf (a, b, c) is encoded using the constructed code table. In step S7, it is determined whether or not all pixels to be encoded have been encoded. If all the pixels have not been encoded, the process returns to step S1 and the subsequent processing is repeated. If all the pixels have been encoded, the encoding process ends.

  FIG. 13 is a diagram illustrating an example of the configuration of the decoding device. A decoding device 80 illustrated in FIG. 13 includes a code table selection unit 81, a pixel value generation unit 82, and a residual decoding unit 83. The encoded data storage memory 86 stores data encoded by the encoding device 30 of FIG. The decoding device 80 receives the encoded data from the encoded data storage memory 86, obtains the pixel value of interest by executing the decoding process, and stores the data of the obtained pixel value in the decoded data storage memory 85. . The data in the decoded data storage memory 85 is supplied to the code table selection unit 81 to be used for context calculation in the decoding process of subsequent pixels in the raster scan order.

  The code table selection unit 81 calculates the first variance value and the second variance value from the peripheral pixel values around the pixel of interest, and selects the code table corresponding to the left-right asymmetric probability distribution having different variance values on the left and right. Selection is made according to the first variance value and the second variance value. The residual decoding unit 83 decodes the residual obtained by subtracting the predicted value from the pixel value of interest using the code table selected by the code table selecting unit 81. That is, the residual decoding unit 83 decodes the residual from the decoding target code by taking out the residual corresponding to the code in the encoded data from the code table. The pixel value generation unit 82 calculates a target pixel value by adding a predicted value to the decoded residual. The code table selection unit 81 includes a context C1 derivation unit 91, a predicted value generation unit 92, an adjacent point selection unit 93, a context C2 derivation unit 94, a first table selection unit 95, a second table selection unit 96, and a code table construction unit. 97. The configuration and operation of the code table selection unit 81 are the same as the configuration and operation of the code table selection unit 31. As described above, when the target pixel is obtained by decoding, that is, when the decoding process is applied to the decoding target code, the pixel values used for obtaining the contexts C1 and C2 are already obtained by the decoding process in the raster scan order. Yes. Accordingly, contexts C1 and C2 can be obtained in the same manner as in the encoding process, and a code table can be selected as appropriate.

  FIG. 14 is a flowchart showing the flow of the encoding process performed by the decoding apparatus 80 shown in FIG.

  In step S1, a context C1 is derived. In step S2, the code table Table1 is determined based on the context C1. In step S3, the predicted value calculation method f (a, b, c) is selected based on the context C1. In step S4, the context C2 is calculated based on the predicted value and the adjacent pixel value, and the code table Table2 is determined based on the context C2. In step S5, a code table corresponding to a left-right asymmetric probability distribution is constructed by combining the code table Table1 and the code table Table2. In step S6, a residual is calculated by applying a decoding process to the decoding target code using the constructed code table. That is, the residual corresponding to the decoding target code is extracted from the code table. In step S7, the pixel value of interest is calculated based on the residual (prediction error) and the predicted value f (a, b, c). In step S8, it is determined whether or not all pixels to be decoded have been decoded. If all the pixels have not been decoded, the process returns to step S1 and the subsequent processing is repeated. If all the pixels have been decoded, the decoding process ends.

  FIG. 15 is a diagram illustrating an example of the configuration of an image processing system that performs encoding and decoding. The image processing system of FIG. 15 includes a digital camera 100, a personal computer (PC) 101, and a display unit 102 such as a CRT or a liquid crystal display device. The digital camera 100 includes an image sensor 111, a volatile memory 112, an encoding unit 113, a nonvolatile memory 114 such as a flash memory, a decoding unit 115, and a display unit 116 such as a liquid crystal display device. The image of the subject 103 captured from the image sensor 111 is temporarily stored in a high-speed volatile memory 112 such as an SRAM (Static Random Access Memory) or a DRAM (Dynamic Random Access Memory). The encoding unit 113 may have the same configuration as the encoding device 30 shown in FIG. 5, encodes and compresses image data read from the volatile memory 112, and converts the encoded image data to a memory card or the like. Store in the nonvolatile memory 114. When displaying the captured image on the display unit 116 of the digital camera 100, the decoding unit 115 built in the camera decodes the image data in the nonvolatile memory 114 and supplies the decoded image data to the display unit 116. The decoding unit 115 may have the same configuration as that of the decoding device 80 shown in FIG. When displaying an image captured by the digital camera 100 on the personal computer 101, the image data stored in the non-volatile memory 114 is taken into the personal computer 101 and decoded by the decoding software 117 in the personal computer 101. At this time, the decoding software 117 may execute the decoding process shown in FIG. The decoded image data is supplied to the display unit 102 and displayed.

  In the above system configuration, more image data can be stored in the nonvolatile memory of the camera by a highly efficient compression method using a code table corresponding to a left-right asymmetric probability distribution. In addition, since the amount of data when stored in a low-speed nonvolatile memory that tends to be a bottleneck for improving the continuous shooting speed can be reduced, the writing time can be shortened.

  As mentioned above, although this invention was demonstrated based on the Example, this invention is not limited to the said Example, A various deformation | transformation is possible within the range as described in a claim.

The present invention includes the following contents.
(Appendix 1)
A first variance value and a second variance value are calculated from the neighboring pixel values around the pixel of interest, and a code table corresponding to a left-right asymmetric probability distribution having different variance values on the left and right is expressed as the first variance value. A code table selection unit that selects according to the second variance value;
An image processing apparatus comprising: an encoding or decoding unit that encodes or decodes a residual obtained by subtracting a predicted value from a pixel value of interest using the selected code table.
(Appendix 2)
The code table selection unit includes a plurality of code tables corresponding to a plurality of probability distributions having a plurality of different variance values, and selects a first code table from the plurality of code tables according to the first variance value The second code table is selected from the plurality of code tables according to the second variance value, and the first code table and the second code table are combined. Image processing apparatus.
(Appendix 3)
The code table selection unit calculates the first variance value according to the surrounding pixel value, and calculates the second variance value according to the surrounding pixel value, an adjacent pixel value, and the predicted value. The image processing apparatus according to appendix 1 or 2, characterized by calculating.
(Appendix 4)
The encoding or decoding unit, when encoding, if the target pixel value is equal to the adjacent pixel value, the residual in the first code table or the second code table corresponds to zero The image processing apparatus according to claim 3, wherein a code is assigned to the residual.
(Appendix 5)
The probability distribution corresponding to the selected code table is such that when the adjacent pixel value is larger than the predicted value, the variance value in the positive region is larger than the variance value in the negative region. Further, when the adjacent pixel value is small, the variance value in the negative region is larger than the variance value in the positive region.
(Appendix 6)
Calculating a first variance value and a second variance value from neighboring pixel values around the pixel of interest;
A code table corresponding to a left-right asymmetric probability distribution having different variance values on the left and right is selected according to the first variance value and the second variance value;
An image processing method comprising the steps of encoding or decoding a residual obtained by subtracting a predicted value from a pixel value of interest using the selected code table.
(Appendix 7)
The selecting step selects a first code table according to the first variance value from a plurality of code tables corresponding to a plurality of probability distributions having a plurality of different variance values, and selects the first code table from the plurality of code tables. The image processing method according to appendix 6, further comprising the steps of selecting a second code table according to a second variance value and combining the first code table and the second code table.
(Appendix 8)
In the calculating step, the first variance value is calculated according to the surrounding pixel value, and the second variance value is calculated according to the surrounding pixel value, the adjacent pixel value, and the predicted value. The image processing method according to appendix 6 or 7, wherein:
(Appendix 9)
In the encoding or decoding step, when the pixel value of interest is equal to the adjacent pixel value at the time of encoding, the residual in the first code table or the second code table corresponds to zero 9. The image processing method according to appendix 8, wherein a code to be assigned is assigned to the residual.

30 Coding Device 31 Code Table Selecting Unit 32 Residual Generating Unit 33 Encoding Unit 41 Context C1 Deriving Unit 42 Predicted Value Generating Unit 43 Adjacent Point Selecting Unit 44 Context C2 Deriving Unit 45 First Table Selecting Unit 46 Second Table Selecting Unit 47 Code table construction part

Claims (4)

Calculates a first dispersion value in accordance with the surrounding pixel value near the target pixel, the second variance value is calculated in accordance with the predicted value of the pixel of interest is calculated on the basis of the neighboring pixel value, the horizontal axis Is a prediction error that is the difference between the value of the pixel of interest and the prediction value, and the vertical axis represents the probability distribution of the occurrence of the prediction error, the first variance value is determined on the left and right sides. a code table selecting unit a code table corresponding to the asymmetric probability distribution having a second variance value on the other hand, is selected according to the first variance and the second variance value has,
The image processing apparatus characterized by comprising an encoding or decoding unit for coding or decoding by the selected code table residual obtained by subtracting the prediction value from the value of the target pixel.   The code table selection unit includes a plurality of code tables corresponding to a plurality of probability distributions having a plurality of different variance values, and selects a first code table from the plurality of code tables according to the first variance value The second code table is selected from the plurality of code tables according to the second variance value, and the first code table and the second code table are combined. The image processing apparatus described. The selected probability corresponding to the code table distribution, wherein when than the predicted value is larger neighbor pixel value is greater than the dispersion value in the region dispersion value is negative at the positive region, than the predicted value 3. The image processing apparatus according to claim 1, wherein when the adjacent pixel value is small, a variance value in a negative region is larger than a variance value in a positive region. The image processing apparatus calculates a first variance value according to the peripheral pixel value around the target pixel, and calculates a second variance value according to the predicted value of the target pixel calculated based on the peripheral pixel value. Calculate
When the image processing apparatus considers a probability distribution in which the horizontal axis is a difference between the value of the target pixel and the predicted value, and the vertical axis is the appearance probability of the prediction error, a code table corresponding to the asymmetric probability distribution having a second variance value on the other hand have the first dispersion value, selected in accordance with the first dispersion value and the second dispersion value ,
Image processing method characterized in that the image processing apparatus, including the step of encoding or decoding by the selected code table residual obtained by subtracting the prediction value from the value of the target pixel.

Images