JP2019106173A - Image processing method, image processing apparatus and program - Google Patents

Abstract

To provide an image processing method, an image processing apparatus, and a program that are capable of realizing patch-based noise reduction processing in which generation of artifact noise is suppressed in an image processing apparatus that generates an output image by reducing noise with respect to an input image.SOLUTION: The image processing apparatus includes: patch of interest setting means that sets a patch of interest having a first number of pixels in an input image; detection means that detects a plurality of similar patches similar to the patch of interest in the input image; target patch setting means that, based on the plurality of similar patches, sets a plurality of target patches including at least one pixel included in the similar patch and having a second number of pixels different from the first number of pixels; noise reduction processing means that performs noise reduction processing on a target patch based on the plurality of target patches; and synthesis means that generates an output image by combining the target patch whose noise has been reduced by the noise reduction processing means.SELECTED DRAWING: Figure 2

Description

  The present invention relates to an image processing technique for reducing noise contained in an input image.

  BACKGROUND Conventionally, noise reduction processing for reducing noise with respect to an image included in an image obtained by capturing with an imaging device is known. Patent Document 1 discloses a method of generating a patch set from a photographed image, performing noise reduction processing on all patches belonging to the patch set, and combining the noise-reduced patches.

JP 2013-026669 A

  In such a conventional patch-based noise reduction method, a plurality of similar patches similar to the target patch to be processed in the captured image are detected as a patch set. At this time, if the number of pixels of the patch of interest is large, the detection accuracy of the similar patch is high. On the other hand, in the noise reduction process for the patch set, when the number of pixels of each patch is larger than the number of patches included in the patch set, the noise reduction process may not be performed accurately for each patch. An object of the present invention is to appropriately execute detection of similar patches and noise reduction processing of each patch in patch-based noise reduction processing.

  In order to solve the above problems, the present invention is an image processing apparatus that generates an output image by reducing noise with respect to an input image, and sets a patch of interest having a first number of pixels in the input image. And at least one pixel included in the similar patch based on the plurality of similar patches, and detection means for detecting a plurality of similar patches similar to the target patch in the input image. Noise for executing noise reduction processing on the target patch based on target patch setting means for setting a plurality of target patches having a second number of pixels different from the first number of pixels, and the plurality of target patches It has a reduction processing means, and a synthesis means for generating the output image by synthesizing the target patch whose noise has been reduced by the noise reduction processing means. It is characterized in.

  According to the present invention, in patch-based noise reduction processing, detection of similar patches and noise reduction processing of each patch can be appropriately performed.

Block diagram showing the hardware configuration of the image processing apparatus Block diagram showing the detailed logical configuration of the image processing apparatus Flow chart showing the flow of noise reduction processing executed by the image processing apparatus A diagram that explains the pixel of interest, the patch of interest, the reference patch, and the target patch Block diagram showing the detailed logical configuration of the image processing apparatus Flow chart showing the flow of noise reduction processing executed by the image processing apparatus Diagram explaining the positional relationship between similar patches and target patches

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. In addition, the structure shown in the following embodiment is only an example, and this invention is not necessarily limited to the illustrated structure.

First Embodiment
In the present embodiment, an image processing apparatus that executes patch-based noise reduction processing will be described. The patch-based noise reduction process includes the steps of detecting similar patches that are similar to the patch of interest, and reducing the noise of each patch using the detected similar patches. In this embodiment, in particular, the number of pixels forming the patch used when detecting similar patches and the number of pixels forming the patch used when reducing noise of each patch are appropriately set. FIG. 1 is a block diagram showing the hardware configuration of the image processing apparatus according to the present embodiment. The image processing apparatus includes a CPU 101, a RAM 102, a storage unit 103, a general-purpose I / F (interface) 110, and a display unit 108. The respective components are connected to one another via a main bus 108. The image processing apparatus is connected to the imaging device 105, the input device 106, and the external memory 107 via the general-purpose I / F 110. The input device 106 is, for example, a device such as a mouse or a keyboard for the user to instruct the image processing apparatus.

  The CPU 101 controls each component in accordance with the input image and a program described later. The storage unit 103 is a storage device such as an HDD or an SSD. The storage unit 103 stores computer programs for the CPU 101 to execute various configurations. The RAM 102 is used as a buffer memory that temporarily holds input image data and the like, a work area of the CPU 101, and the like. The image processing apparatus causes the CPU 101 to interpret the program stored in the storage unit 103, and performs an operation based on an instruction. The display unit 108 is a liquid crystal display, an organic EL display, or the like. The display unit 108 displays a user interface (UI) for inputting an image or a desired instruction by the user. The image processing apparatus executes noise reduction processing on the image stored in the RAM 102 in accordance with an instruction from the user input via the display unit 108. Further, the image after the noise reduction processing is stored again in the RAM 102. The image after noise reduction processing stored in the RAM 102 is output to the display unit 108 or the external memory 107 according to an instruction from the user.

  Here, an outline of patch-based noise reduction processing will be described. A patch formed of a plurality of pixels with respect to the target pixel in the input image is referred to as a target patch. Each of a plurality of pixels in the vicinity of the target pixel is referred to as a reference pixel, and a patch corresponding to each reference pixel is referred to as a reference patch. The reference patch is a patch composed of a plurality of pixels based on the reference pixel. The reference patch is referred to when performing noise reduction processing on the patch of interest. Since a plurality of reference pixels are set for one target pixel, a plurality of reference patches exist for one target pixel. The patch-based noise reduction process detects similar patches which are patches having similar distribution of pixel values to the focused patch among the reference patches, and generates a similar patch group (patch set). The similar patch is a patch configured with the same number of pixels as the patch of interest. The similar patch group may include the patch of interest itself. Next, based on the similar patch group, noise reduction processing is performed to reduce the noise of each patch included in the similar patch group. By combining similar patches with reduced noise, an image reduced in noise with respect to the input image is generated. The patch means an area corresponding to a part of the input image. Moreover, a patch means the pixel group comprised by several pixel. FIG. 2 is a block diagram showing a detailed logical configuration of the image processing apparatus that executes the noise reduction processing. The image processing apparatus includes an image input unit 201, a patch of interest setting unit 202, a similar patch detection unit 203, a target patch setting unit 204, a noise reduction processing unit 205, and an image combining unit 206. The image input unit 201 inputs image data representing an input image to be subjected to noise reduction processing. The input image is an image of each of R (red), G (green) and B (blue). The image corresponding to each color consists of pixels having 8-bit pixel values of 0-255.

  The focused patch setting unit 202 sets a focused patch in the input image. The shape of the patch of interest is preset. In the present embodiment, pixels in the input image are set as a target pixel in raster order, and a rectangular area of 8 pixels × 8 pixels including the target pixel is set as a target patch.

  The similar patch detection unit 203 detects a similar patch by setting a plurality of reference patches in the vicinity of the target pixel and selecting a patch similar to the target patch among the reference patches as a similar patch. In the present embodiment, the similar patch detection unit 203 selects a plurality of reference patches as similar patches. Specifically, the similar patch detection unit 203 first sets, as reference pixels, pixels in the vicinity of the target pixel included in the predetermined range in order, and sets a reference patch for the reference pixel in the same manner as the target patch. Next, the similar patch detection unit 203 derives the similarity between the reference patch and the patch of interest, and determines whether each reference patch is similar to the patch of interest. When the reference patch is similar to the patch of interest, the reference patch is selected as a similar patch. Here, 15 pixels × 15 pixels centered on the pixel of interest are set as the predetermined range in which the reference pixels are set. Therefore, the similar patch detection unit 203 sets 225 pixels including the target pixel as a reference pixel, and detects up to 225 similar patches. The similar patch to be detected is the same size as the patch of interest. Therefore, the similar patch in the present embodiment is a rectangular patch consisting of 8 pixels × 8 pixels.

  The target patch setting unit 204 sets a target patch to be used for noise reduction by adjusting the number of pixels of the detected similar patch. Although the details will be described later, as the number of pixels of the patch of interest is large, high-accuracy noise reduction can not be realized unless the subsequent processing is performed using a large number of similar patches. On the other hand, if the number of pixels of the patch of interest is reduced for highly accurate noise reduction processing, the accuracy of detecting a patch similar to the patch of interest will be degraded. Therefore, in the present embodiment, the target patch setting unit 204 sets a target patch by adjusting a similar patch consisting of 8 pixels × 8 pixels to a patch consisting of 4 pixels × 4 pixels. As described above, in the present embodiment, the number of pixels of patches used when detecting similar patches and the number of pixels of patches used when performing noise reduction processing of each patch are set using different numbers of pixels. Here, the target patch is set for every detected similar patch. Here, since a part of similar patches is a target patch, the target patch group can also be regarded as a patch similar to each other.

  The noise reduction unit 205 performs noise reduction on each target patch based on the target patch group, and generates a noise-reduced target patch group. The image combining unit 206 combines the noise-reduced target patch group to generate an output image subjected to noise reduction processing. The image combining unit 206 places the similar patch subjected to noise reduction processing at the original position in the input image, and in the case of overlapping with another corrected similar patch, a value obtained by averaging the values returned to the same pixel position. Calculate The noise-reduced target patch group is synthesized by updating the pixel value of the original pixel with the pixel value obtained as a result of the noise reduction processing or the averaged value. Such combination processing is also called aggregation.

  Subsequently, a specific process flow of the noise reduction process performed by the above-described image processing apparatus will be described. FIG. 3 is a flowchart showing the flow of noise reduction processing according to the present embodiment. The CPU 101 reads and executes a program for realizing the flowchart shown in FIG. In addition, each process (step) in a flowchart is represented using "S".

  In step S301, the image input unit 201 acquires an image. As described above, images corresponding to each of RGB are sequentially input. The subsequent processing is performed on each color image. In S302, the patch of interest setting unit 202 sets a patch of interest consisting of pixels of the first number of pixels in the acquired image. As described above, in the present embodiment, an area of 8 pixels × 8 pixels is used as a patch of interest. FIG. 4A shows the positional relationship between the pixel of interest and the patch of interest in the present embodiment. One rectangle represents a pixel. In FIG. 4A, the blackened pixel is the target pixel. A rectangle with four corners of (x-3, y-3), (x + 4, y-4), (x-3, y + 4), (x + 4, y + 4), where (x, y) is the pixel position of the pixel of interest An area of 64 pixels included in is set as a target patch.

  In S303, the similar patch detection unit 203 detects a patch similar to the patch of interest based on the patch of interest. Specifically, first, the similar patch detection unit 203 sequentially sets the pixels included in the search range in the vicinity of the pixel of interest as reference pixels, and sets reference patches for the reference pixels. The shape of each reference patch matches the shape of the patch of interest. In the present embodiment, 15 pixels × 15 pixels centered on the target pixel are set as a search range for similar patches. For example, assuming that the pixel position of the pixel of interest is (x, y) and the (x-3, y-3) pixel is a reference pixel, 8 pixels including the reference pixel as shown in FIG. An 8-pixel area is set as a reference patch. Based on the similarity between the focused patch and the reference patch, it is determined whether the referenced patch is similar to the focused patch. The similarity is calculated by calculating the difference sum of squares obtained by combining the squares of the differences between the pixels in the patch of interest and the pixels in the reference patch for each pixel and the pixel values of corresponding pixels. Ru. Specifically, the similarity patch detection unit 203 calculates the difference square sum SSDi of the i-th reference patch according to the following equation (1).

  Ri (j) is the j-th pixel value in the i-th reference patch. T (j) is the j-th pixel value in the patch of interest T. The sum of squared differences is calculated by accumulating a value obtained by squaring the difference between the pixel values of two pixels having the same pixel position in the patch. As described above, when the similarity is calculated using the difference square sum SSDi, the smaller the value of the similarity, the more the reference patch Ri is similar to the patch of interest T. That is, the smaller the similarity value, the higher the similarity. On the other hand, the reference patch Ri is not similar to the patch of interest T as the similarity value is larger.

  If the similarity value of the reference patch is less than a predetermined threshold value, the similar patch detection unit 203 determines that the patch is similar to the focused patch, and sets the reference patch as a similar patch. On the other hand, if the similarity is equal to or greater than a predetermined threshold, it is determined that the reference patch is not similar to the patch of interest. Thereby, the similar patch detection unit 203 detects similar patches similar to the feature of the focused patch in the vicinity of the focused pixel. Here, since the search range is 15 pixels × 15 pixels, the similarity is determined 225 times for one patch of interest. It is assumed that the similar patch detection unit 203 detects N similar patches. The number of similar patches depends on whether each reference patch is similar to the patch of interest. Therefore, in many cases, the number of similar patches is different when the focused patch is different.

  In the next S304, the target patch setting unit 204 sets a target patch for performing the noise reduction processing based on the similar patch group. The target patch is configured by pixels of a second pixel number different from the first pixel number for detecting a similar patch. FIG. 4A shows the relationship between the position of the similar patch and the target patch detected in the present embodiment and the number of pixels. The target patch is a patch set corresponding to the similar patch, and the number of pixels different from the number of pixels of the focused patch. In the present embodiment, as described above, when the black pixel is a pixel of interest, the target patch setting unit 204 sets an area of 4 pixels × 4 pixels including the pixel of interest and the hatched pixel as a target patch. That is, in the present embodiment, the target patch is a partial area of a patch similar to the patch of interest. Thus, the target patch setting unit 204 sets target patches for all similar patches detected by the similar patch detection unit 203. The number of target patches is N, similar to the number of similar patches.

  In S305, the noise reduction processing unit 205 performs noise reduction processing of each target patch based on the target patch group, and generates a noise-reduced patch group. First, the noise reduction processing unit 205 calculates the average patch and the covariance matrix C based on the target patch group. In the target patch group, an average of pixel values of pixels at the same position is calculated, and an average value is stored at each pixel position to calculate an average patch. Thus, the average patch has the same shape as the shape of the target patch. In the present embodiment, it is a patch of 4 pixels × 4 pixels. Assuming that the number of pixels in the average patch is M and the column vector in which the pixel values of the M pixels are arranged is Q, the noise reduction processing unit 205 calculates the average patch by Equation (2).

  Pi represents the i-th target patch. The covariance matrix C calculated from the target patch group is a square matrix, and the size of one side is the number M of pixels forming the target patch. The noise reduction processing unit 205 calculates the covariance matrix C by the equation (3) using the average patch Q.

  The covariance matrix C indicates how there is a correlation between each pixel in the target patch. When focusing on two pixels adjacent to each other in a patch, if the pixel values are close to each other in many target patches or if there is a similar difference, it can be regarded that there is a correlation between the two pixels.

  Furthermore, the noise reduction processing unit 205 calculates an eigenvector corresponding to each of a plurality of eigenvalues and the eigenvalues of the covariance matrix C. Texture components are extracted from the target patch group by calculating eigenvalues and eigenvectors for a covariance matrix indicating a correlation between pixels. In general, a plurality of eigenvalues and eigenvectors exist, and the number of each of the eigenvalues and eigenvectors is the size of one side of the covariance matrix. That is, there are M eigenvalues and eigenvectors for the number of pixels constituting the target patch. Further, the number of elements of the eigenvector is the same as the number M of pixels constituting the target patch. The j-th eigenvalue is denoted by λ j and the j-th eigenvector is denoted by E j. Eigenvalues λj and eigenvectors Ej of covariance matrix C satisfy the relation of equation (4) for covariance matrix C. That is, the eigenvalue λj and the eigenvector Ej correspond to each other.

CE _j = λ _j E _j (4)
The noise reduction processing unit 205 calculates the eigenvalue λj and the eigenvector Ej which satisfy the equation (4). Here, the eigenvectors Ej correspond to texture components extracted in patch sets similar to one another. In addition, the eigenvalue λj is approximately the variance of noise in the image when the texture component represented by the corresponding eigenvector is a texture caused by noise. As the value of the eigenvalue is larger than the variance of the noise, it means that the eigenvector corresponding to the eigenvalue is a texture component of the subject which should be left in the image. The noise reduction processing unit 205 performs noise reduction processing on each target patch so as to leave the texture component of the subject and remove the texture component due to noise.

  The noise reduction processing unit 205 generates a basis matrix B corresponding to each target patch according to Expression (5) based on the eigenvalues and the eigenvectors.

B ≡ (E ₁ E ₂ · · E _M ) (5)
The basis matrix B is a matrix of K rows and M columns. Furthermore, the noise reduction processing 205 performs projection processing on each target patch based on the basis matrix B. The coefficient K is set in advance. The coefficient K saw 0 <K <M and corresponds to the number of principal components left as a texture. The noise reduction processing unit 205 derives the projection matrix H from the basis matrix B by equation (6).

H ≡ (E ₁ E ₂ · · E _K ) ^t (E ₁ E ₂ · · E _K ) (6)
The noise reduction processing unit 205 projects the difference patch obtained by subtracting the average patch W from each target patch Pi as in equation (7) using the projection matrix H, and projects the projection using the projection matrix H. By adding to the patch, the noise of the target patch is reduced. By such correction, the patch Oi reduced in noise with only the main component left is calculated. The process by the noise reduction process 205 is expressed as the following equation (7).

O _i ≡ Q + H (P _i- Q) (7)
The noise reduction processing unit 205 accumulates the pixel values of the pixels included in each target patch subjected to noise reduction.

  In S306, the patch of interest setting unit 202 determines whether the patch of interest has been set for each pixel in the input image. If the target patch is set with all pixels as the target pixel, the process advances to step S307. If there is a pixel that is not set as a target pixel, the process returns to S302 to repeat the processing.

  In S307, the image combining unit 206 combines the noise-reduced patch groups (aggregation) to generate an output image in which noise in the input image is reduced. Specifically, in an image configured to have the same size as the input image, the pixel values of the respective pixels included in the target patch after the noise reduction processing are accumulated at the pixel position where the patch before noise reduction was present. However, when a plurality of pixel values after noise reduction processing are calculated, the pixel values are updated using the average value of the plurality of pixel values.

  Thus, the noise reduction process according to the present embodiment is completed. In the present embodiment, in the process of detecting a similar patch similar to the patch of interest, an area of 8 pixels × 8 pixels is used as the patch size. Thereafter, in the process of performing noise reduction processing on the patch, a partial region of a similar patch consisting of 8 pixels × 8 pixels is set as a target patch. In the process of detecting similar patches, if the patch size is small, detection accuracy of similar patches is likely to be low. Even if the reference patch is not similar to the patch of interest, the similarity value is calculated to be small. Therefore, when detecting similar patches, similar patches can be detected with higher accuracy when the number of pixels of the patch is larger than when the number of pixels of the patch is small.

  On the other hand, the eigenvalues and eigenvectors calculated from the patch group correspond to the texture components extracted from the patch as described above. Eigenvalues and eigenvectors are calculated for the number of pixels constituting the patch. Thus, the larger the size of the patch, the more texture components will be extracted. However, when the number of patches is small, it is difficult to extract many patch-specific texture components. In other words, when the number of patches is fixed, the smaller the texture component (eigen vector) to be extracted, the better the accuracy. Therefore, in the present embodiment, in the process of calculating eigenvalues and eigenvectors for reducing the noise of a patch, a patch having a smaller number of pixels than similar patches is used as a target patch to reduce the number of eigenvalues and eigenvectors to be calculated. . Thereby, the eigenvalues and eigenvectors used for the noise reduction process can be calculated with high accuracy. As a result, appropriate noise reduction processing can be performed on the patch group, and as a result, noise in the input image can be further reduced.

  In the present embodiment, aggregation is performed collectively after setting all pixels as the target pixel. However, the process of combining patches that have undergone noise reduction processing so far at a predetermined timing may be repeated on the input image. For example, a pixel included in a predetermined band area in the input image may be set as a pixel of interest, and the aggregation may be performed when processing for all the pixels in the band area is completed. Alternatively, when a predetermined number of pixel values after noise reduction processing are obtained, averaging may be performed, and the average value may be sequentially output from the calculated pixels.

  Further, in the present embodiment, both the similar patch and the target patch are rectangular areas. Besides, a circular area or a rod-like area may be used as a patch. Also, a pixel group including pixels other than adjacent pixels may be used as a patch. Further, in the present embodiment, the target patch is set to be a partial region in the similar patch. Therefore, the target patch is an area that can be regarded as similar to the patch of interest. However, for example, as shown in FIG. 4C, even when a part of the target patch is not included in the similar patch, an effect can be obtained. If the similar patch is similar to the patch of interest, it is highly likely that the periphery of the similar patch and the periphery of the patch of interest are also similar. Therefore, even if the target patch is set as shown in FIG. 4C, the target patch group can be regarded as sufficiently similar to each other. Also, in some cases, as shown in FIG. 4D, it is desirable to set a patch larger than the similar patch as a target patch. For example, when a large number of similar patches are detected, a large number of texture components (eigenvectors) can be accurately calculated even if the number of pixels of the target patch is increased. On the other hand, when many similar patches are detected, if the number of pixels of the patch of interest is increased and the similar patches are detected again, the number of patches determined to be similar decreases. In this case, the noise reduction effect can be further enhanced by increasing the number of pixels of the target patch more than the number of pixels of the similar patch.

Second Embodiment
In the first embodiment, the method of setting the target patch having the number of pixels different from the number of pixels of the similar patch by adjusting the size of the similar patch has been described. In the second embodiment, a method of switching between the case where the size of the similar patch is not adjusted and the case where the size of the similar patch is adjusted in the noise reduction processing for the patch will be described. The hardware configuration of the image processing apparatus is the same as that of the first embodiment. About the structure similar to Embodiment 1, the same code | symbol is attached | subjected and detailed description is abbreviate | omitted.

  FIG. 5 is a block diagram showing a detailed logical configuration of the image processing apparatus in the second embodiment. The image processing apparatus includes an image input unit 201, a patch of interest setting unit 202, a similar patch detection unit 203, a determination unit 501, a target patch setting unit 204, a noise reduction processing unit 205, and an image combining unit 206.

  When performing noise reduction processing on a patch group, the determination unit 501 determines whether to set a target patch having the number of pixels different from that of the similar patch. When the target patch is not set, noise reduction processing is performed using a plurality of detected similar patches. This can be reworded as using similar patches as target patches as they are. In the present embodiment, the determination unit 501 determines whether or not to set the target patch based on the variance of pixel values in the range in which the reference pixel is set when detecting the similar patch and the noise variance in the input image. . The noise variance in the input image is calculated in advance based on the input image.

  The noise reduction processing unit 502 executes noise reduction processing on each target patch based on the target patch group when the target patch is set. Further, when it is determined that the target patch is not set by the determination unit, the noise reduction processing unit 502 performs noise reduction processing on each similar patch based on the similar patch group.

  FIG. 6 is a flowchart showing the flow of noise reduction processing in the second embodiment. S301, S302, S303, S304, and S306 are the same as in the first embodiment.

  In step S601, the determination unit 501 determines whether to set a target patch. The determination unit 501 determines whether to set the target patch based on the variance of pixel values within the range in which the reference pixels are set. First, the determination unit 501 calculates the variance of pixel values based on the pixel value of each pixel in the search range (15 pixels × 15 pixels here) in which the reference pixel is set for the pixel of interest. If the variance of the pixel values of each pixel in the search range is larger than the noise variance of the input image, the determination unit 501 determines to set a target patch with a similar patch and a different number of pixels. If the variance of pixel values in the search range is smaller than the noise variance of the input image, target patches with different numbers of similar patches and pixels are not set.

  When it is determined by the determination unit 501 that the target patch is to be set, a partial area of the similar patch is set as the target patch. If there is a texture, the variance of pixel values may be large, and the number of similar patches that can be detected may be small. When the number of similar patches is small, if eigenvectors and eigenvectors similar to those of the first embodiment are calculated based on the similar patch group with the large number of pixels, similar eigenvectors corresponding to texture components can not be calculated appropriately. Therefore, in the present embodiment, in order to realize the noise reduction processing with high accuracy even if the number of similar patches detected is small, a target patch having a smaller number of pixels than the number of pixels of similar patches is set. On the other hand, if the variance of pixel values is smaller than the noise variance, the search area is flat and there is a high possibility that a large number of similar patches can be detected. When a large number of similar patches can be detected, the noise reduction effect is higher if the noise reduction processing is performed using a patch having a large number of pixels than the noise reduction processing using a patch having a small number of pixels. Therefore, when the variance of the pixel values is sufficiently small, the target patch of the number of pixels smaller than the number of pixels of the similar patch is not set, and the noise reduction processing is performed using the similar patch.

  In S602, the noise reduction processing unit 502 performs noise reduction processing based on the target patch when the target patch is set, and based on the similar patch when the target patch is not set. The noise reduction process in S602 is similar to S305 in the first embodiment. Thus, the noise reduction process in the second embodiment is completed. In this way, noise reduction processing is performed according to the feature of the image by switching the number of pixels of the patch for detecting similar patches and the number of pixels of the patch for noise reduction processing for the patches to be different. be able to. When the number of similar patches is large, noise reduction processing is enhanced by performing noise reduction processing using patches with a large number of pixels, and when the number of similar patches is small, noise reduction processing is performed using patches with a small number of pixels. Noise can be reduced with high accuracy.

<Modification of Embodiment 2>
In the second embodiment, when the variance of pixel values is large, a target patch having a smaller number of pixels than the number of pixels of similar patches is set. When the variance of pixel values is small, the target patch is not set. On the other hand, the noise reduction process using the similar patch and the noise reduction process using the target patch having a larger number of pixels than the similar patch may be switched. In this case, when the variance of pixel values is smaller than the noise variance, the determination unit 501 determines to set a target patch having a larger number of pixels than a similar patch. If the variance of the pixel values is equal to or greater than the noise variance, the determination unit 501 does not set the target patch. In a region where similar patches can be made, performing noise reduction processing using a larger patch can prevent low frequency noise from remaining as a result of the noise reduction processing. The determination unit 501 determines whether the patch size (number of pixels) used to detect similar patches should be different from the patch size (number of pixels) when performing the noise reduction processing. Patches of appropriate size can be set in the process.

  In the second embodiment, the determination unit 501 determines whether to change the size of the patch used for the processing based on the variance of the pixel values of the pixels included in the search range. However, as a range for calculating the variance of pixel values, a group of pixels included in the focused patch and the similar patch group may be used.

  When the number of similar patches is large, the number of pixels of patches used for noise reduction processing is large, and when the number of similar patches is small, the number of pixels of patches used for noise reduction processing is reduced. In the second embodiment, the variance of the pixel values is calculated, and it is determined whether or not the vicinity of the pixel of interest is a flat area where a similar patch can be easily detected. As another example, the determination unit 501 may determine whether to set the target patch using the number of similar patches or the total number of pixels of the similar patch group. In this case, when the number of similar patches or the total number of pixels in the similar patch group is small, the same effect as that of the second embodiment can be obtained by setting patches having a smaller number of pixels than similar patches as target patches. it can. In addition, if it is a flat area, there is a high possibility that a large number of similar patches can be detected, and if it is a non-flat area, a large number of similar patches can not be detected. Therefore, an edge in a focused patch or a search range may be detected, and the determination unit 501 may determine whether to set the target patch according to the edge.

  Furthermore, the determination unit 501 may be configured to specify the number of pixels and the shape of the target patch as well as whether or not to set the target patch. For example, the determination unit 501 may derive the number of pixels according to the dispersion of pixel values in the vicinity of the pixel of interest, and set a target patch of the derived number of pixels.

Embodiment 3
In the first and second embodiments, as a noise reduction process for a patch group, a method of using a projection matrix has been described as an example in which a projection matrix is derived from eigenvalues and eigenvectors of a covariance matrix calculated from the patch group. However, as a patch-based noise reduction process, there is a method of performing a noise reduction process on a patch using a covariance matrix instead of a projection matrix. In the third embodiment, when patch-based noise reduction processing is performed using a covariance matrix, an embodiment in which a target patch is set will be described.

  In the third embodiment, the configuration of the image processing apparatus and the flow of processing are the same as in the first embodiment. However, the processing in S304 executed by the target patch setting unit 202 and the specific processing in S305 executed by the noise reduction processing unit 205 are different.

  First, the details of the process in S305 of the third embodiment will be described. The noise reduction processing 205 in S305 is similar to that of the first embodiment up to the processing of calculating the covariance matrix. Here, the noise reduction processing 205 calculates the correction matrix H based on the covariance matrix C.

H ≡ σ ² (E ₁ E ₂ · · E _M ) ^t diag (λ ₁ ^-1 , λ ₂ ^-1 , · · · λ _M ^-1 ) (E ₁ E ₂ · · E _M ) (8)
The correction matrix H calculated here is, like the covariance matrix C, a square matrix in which the size of one side is the number M of pixels of the target patch. The noise reduction processing unit 205 corrects each of the target patches based on the correction value matrix H. Specifically, when the average patch is Q and the i-th target patch is Pi, the noise reduction processing unit 205 calculates the similar patch Oi corrected by the following equation (6).

O _i ≡P _i -H (P _i -Q) (9)
That is, H (Pi-Q) of the second item in the equation (9) is a correction value for the target patch Pi. By subtracting this correction value from the target patch Pi, the noise of each similar patch is reduced.

  In such noise reduction processing, the inverse matrix of the covariance matrix is used as the correction matrix, as shown in equation (8). The covariance matrix is a square matrix of M × M pixels of the target patch, as described above. In order to calculate a regular matrix (a matrix that can calculate an inverse matrix) as a correction matrix of M × M, the number of patches for the number of pixels + 1 is required. Therefore, when the number of similar patches detected is small, if the covariance matrix is calculated using a similar patch having a large number of pixels, the inverse matrix can not be derived, and as a result, noise reduction processing may not be performed on the patches.

  Therefore, in the third embodiment, the target patch setting unit 202 sets the number of pixels of the target patch according to the number of similar patches detected in S304. Specifically, the target patch setting unit 202 selects, from among the pixels included in the similar patch, the number of minutes of the detected similar patch minus one pixel, and sets it as a target patch. The order of pixels to be selected among similar patches is held in advance. Here, the order is closer to the reference pixel. If (the number of similar patches-1) is larger than the number of pixels of similar patches, all pixels of similar patches are selected. That is, if the similar patch can be detected (the number of pixels of the similar patch + 1) or more, the similar patch is directly used as the target patch.

  When the number of similar patches detected is small, if the number of pixels of similar patches is reduced and similar patches are re-detected, the determination accuracy of similar patches is reduced. Therefore, when the number of similar patches is small, the number of pixels of the target patch is set according to the number of similar patches. As a result, even if the number of similar patches detected is small, noise reduction processing can always be performed on the patch group.

  As in the second embodiment, the third embodiment may be switched to the case where the similar patch is used for the noise reduction processing as it is.

  In the third embodiment, the method of setting the number of pixels of the target patch according to the number of similar patches detected has been described as an example. However, as in the first embodiment, regardless of the detection of the similar patch, a patch having a pixel number different from the pixel number of the similar patch, such as a partial region of the similar patch, may be set as the target patch.

Fourth Embodiment
In the above-described embodiment, target patches having different numbers of pixels from similar patches are set for noise reduction processing on patch groups, and the eigenvalues and eigenvectors of the covariance matrix and the correction matrix are obtained from the patch group consisting of a plurality of target patches. Described how to derive. In the fourth embodiment, after calculating the eigenvalues and eigenvectors of the covariance matrix and the like from the similar patch group, a method of performing noise reduction processing on a target patch having the number of pixels different from that of the similar patch will be described. In the fourth embodiment, the configuration of the image processing apparatus and the flow of processing are the same as in the third embodiment. However, the specific processing in S304 executed by the target patch setting unit 204 and in S305 executed by the noise reduction processing unit 205 are different.

  First, the details of the process in S304 of the present embodiment will be described. In S304, the target patch setting unit 204 sets a target patch based on the similar patch. In the present embodiment, the target patch setting unit 202 acquires defect information indicating a pixel (hereinafter, referred to as a saturated pixel) having a defect such as black crushing or whiteout in an image. For example, the defect information is information in which information indicating whether each pixel is blackout or overexposure is stored. In addition, the information which shows a pixel with crushing or whiteout can be detected by a well-known method. The target patch setting unit 202 refers to the defect information and selects only pixels other than the saturated pixel as the pixels constituting the target patch when the detected similar patch group includes saturated pixels having black crushing or overexposure. Do. However, for any similar patch in the similar patch group, the pixel at the same position in the patch is selected as the target patch. For example, it is assumed that only the upper left pixel of one similar patch in the similar patch group is a saturated pixel. In this case, for each similar patch included in the similar patch group, a pixel group excluding the upper left pixel is set as a target patch.

  If saturated pixels exist in the patch group for which the covariance matrix is to be calculated, the variance may become abnormally small or large, which may cause an artifact to be generated. Therefore, in the present embodiment, the target patch setting unit 202 sets a patch group excluding saturated pixels. Further, in the present embodiment, the target patch is set so as not to be out of the range of the similar patch.

  Next, details of the process in S305 of the fourth embodiment will be described. In S305, the noise reduction processing 205 calculates an average patch of the similar patches and a covariance matrix Cr based on the detected similar patches. In the similar patch group, an average of pixel values of pixels at the same position is calculated, and an average value is stored at each pixel position to calculate an average patch of similar patches. Therefore, the average patch of similar patches has the same shape as the shape of similar patches, and in this embodiment, is a patch of 8 pixels × 8 pixels. Assuming that the number of pixels in the average patch of similar patches is R and the column vector in which the pixel values of R pixels are arranged is Qr, the noise reduction processing unit 205 calculates the average patch of similar patches according to equation (9) Do.

  Si represents the i-th similar patch. The covariance matrix C calculated from the similar patch group is a square matrix, and the size of one side is the number R of pixels constituting the similar patch. The noise reduction processing unit 205 calculates the covariance matrix C by Equation (10) using the average patch Qr of similar patches.

  Here, the noise reduction processing 205 calculates the correction matrix candidate F using Expression (11) based on the covariance matrix Cr.

F ≡ σ ² (E ₁ E ₂ · · E _R ) ^t diag (λ ₁ ^-1 , λ ₂ ^-1 , · · · λ _R ^-1 ) (E ₁ E ₂ · · E _R ) (11)
Similar to the covariance matrix C, the correction matrix candidate F calculated here is a square matrix whose one side size is the number R of pixels of similar patches. Furthermore, the noise reduction processing unit 205 calculates the correction matrix H based on the correction matrix candidate F. The correction matrix H is obtained by extracting a part of the correction matrix candidate F. When the number of pixels of the target patch is M, the correction matrix H is also an M × M square matrix.

  How to extract from the correction matrix candidate F depends on the positional relationship between the similar patch and the target patch. The method of calculating the correction matrix H based on the correction matrix candidate F will be described in detail below. For example, the case where the similar patch and the target patch have the positional relationship of FIG. 4A is taken as an example. FIG. 7 shows column vectors of similar patches. The column vector of similar patches is a vector generated by rearranging the pixels of similar patches in raster order and arranging the pixel values of the respective pixels in one column. Here, when the similar patch and the target patch have the positional relationship of FIG. 4A, the elements that are hatched in FIG. 7 are also included when generating the column vector of the target patch. That is, the column vector of the target patch is included in part of the column vector of the similar patch. When setting the target patch, the target patch setting unit 204 generates, as pixel position information, information indicating which number element among the elements in the column vector of the similar patch has been set as the target patch column vector. , Hold. Here, a plurality of numbers indicating which element is included in the target patch from the top (head) of the column vector of the similar patch is held as the target patch number B. The target patch number B has M elements. When the similar patch and the target patch have a relationship as shown in FIG. 7, when the target patch number B is calculated, “10, 11, 12, 13, 18, 19, 20, 21, 26, 27, 28, 29, 34 , 35, 36, 37 ".

  The noise reduction processing 205 refers to the information indicating the target patch number obtained from the target patch setting unit 205, and extracts only the elements included in the target patch number B in both the row and the column from among the correction matrix candidates F for correction. The matrix H is calculated. For example, when the target patch number B is "10, 11, 12, 13, 18, 19, 20, 21, 26, 27, 28, 29, 34, 35, 36, 37", the correction matrix candidate F is firstly The element in the tenth column and the tenth row is extracted from the first to the first column in the correction matrix H. Subsequently, the eleventh matrix and the eleventh matrix of the correction matrix candidate F are set as the second column and the second row in the correction matrix H. By repeating such a process for the target patch number B, the correction matrix H is generated from the correction matrix candidate F. Further, the noise reduction processing unit 205 extracts only the elements included in the target patch number B from the average patch Qr of similar patches, and calculates the average patch Q corresponding to the target patch. The noise reduction processing unit 205 corrects each of the target patches based on the correction matrix H. Specifically, when the average patch is Q and the i-th target patch is Pi, the noise reduction processing unit 205 calculates a similar patch Oi corrected by the following equation (12).

O _i ≡P _i -H (P _i -Q) (12)
The subsequent processing is the same as that of the third embodiment. If the similar patch group includes saturated pixels, the noise distribution in the similar patch group is significantly reduced, and if the number of patches is not detected relative to the number of patch pixels, artifacts due to overexposure or overexposure may occur. May occur. Therefore, in the present embodiment, when there is a pixel that is subjected to whiteout or black squashing from the similar patch, a target patch excluding those saturated pixels is set. This makes it possible to suppress the occurrence of an artifact and perform the noise reduction processing of each patch with higher accuracy.

<Other Embodiments>
In the above-described three embodiments, software realized by operating a computer program has been described as an example. However, some or all of the components of the block diagrams shown in FIGS. 2 and 5 may be realized by a dedicated image processing circuit.

  Furthermore, the present invention supplies a program that implements one or more functions of the above-described embodiments to a system or apparatus via a network or a storage medium, and one or more processors in a computer of the system or apparatus read and execute the program. Can also be realized.

201 image input unit 202 patch of interest setting unit 204 target patch setting unit 205 noise reduction processing unit 206 image combining unit

Claims (20)

An image processing apparatus that generates an output image by reducing noise with respect to an input image, and a patch of interest setting unit that sets a patch of interest having a first number of pixels in the input image;
Detection means for detecting a plurality of similar patches similar to the patch of interest in the input image;
Target patch setting means for setting a plurality of target patches including at least one pixel included in the similar patch based on the plurality of similar patches and having a second number of pixels different from the first number of pixels ,
Noise reduction processing means for performing noise reduction processing on the target patch based on the plurality of target patches;
An image processing apparatus comprising: combining means for generating the output image by combining target patches whose noise has been reduced by the noise reduction processing means.   The image processing apparatus according to claim 1, wherein the target patch setting unit sets a patch having a smaller number of pixels than the first number of pixels as the target patch.   The image processing apparatus according to claim 1, wherein the target patch setting unit sets a partial area of the similar patch as the target patch. Furthermore, the noise reduction processing means has determination means for determining whether to use a similar patch or a target patch as a patch to be processed.
The noise reduction processing means executes noise reduction processing on the plurality of similar patches based on the plurality of similar patches when it is determined by the determination means that the similar patch is used.
The noise reduction process is performed on the target patch based on the plurality of target patches when it is determined by the determination unit that the target patch is to be used. The image processing apparatus according to claim 1.   The image processing according to claim 4, wherein the determination unit determines whether to use a similar patch or a target patch based on a dispersion of pixel values of pixels included in a region in the input image. apparatus.   The image processing apparatus according to claim 4, wherein the determination unit determines whether to use a similar patch or a target patch based on the number of similar patches detected by the detection unit.   The said determination means detects an edge with respect to the area | region in the said input image, and determines whether it uses a similar patch or an object patch based on the edge in the said area | region, It is characterized by the above-mentioned. Image processing device.   The noise reduction processing means calculates an average patch and a covariance matrix of the plurality of target patches, calculates a basis matrix based on the eigenvalues and the eigenvectors of the covariance matrix, and uses the basis matrix to calculate the plurality of the plurality The image processing apparatus according to any one of claims 1 to 7, wherein noise reduction processing is performed on the plurality of target patches by performing projection processing on the target patch.   The noise reduction processing means calculates an average patch and a covariance matrix of the plurality of target patches, and performs noise reduction processing on the plurality of target patches using an inverse matrix of the covariance matrix. The image processing apparatus according to any one of claims 1 to 7, wherein   10. The target patch setting means acquires defect information indicating whether the pixel is blackout or overexposure for each pixel, and sets the target patch based on the defect information. The image processing apparatus according to any one of the above.   The target patch setting unit sets the plurality of target patches such that there is no pixel indicating that the defect information is black crushing or overexposure in the plurality of target patches. 10. The image processing apparatus according to 10. The target patch setting unit generates pixel position information indicating a pixel position set as the target patch among a group of pixels constituting the similar patch,
The image processing apparatus according to any one of claims 1 to 11, wherein the noise reduction processing unit performs noise reduction processing on the plurality of target patches based on the pixel position information.   The noise reduction processing means calculates a first average patch and a candidate for correction matrix based on the plurality of similar patches, and further, based on the pixel position information, calculates the first average patch and the correction matrix. The second average patch and the correction matrix are calculated from the candidates, and the noise reduction processing of the plurality of target patches is performed using the second average patch and the correction matrix. Image processing apparatus as described.   The noise reduction processing means calculates a first covariance matrix based on the plurality of similar patches, and calculates candidates for the correction matrix using the first covariance matrix. Item 14. An image processing apparatus according to item 13.   The correction matrix candidate is a matrix of a size according to the number of pixels constituting the similar patch, and the correction matrix is a matrix of a size according to the number of pixels constituting the target patch The image processing apparatus according to claim 13 or 14.   The image processing apparatus according to any one of claims 1 to 15, wherein the similar patch and the target patch are rectangular areas which are aligned with a plurality of pixel groups.   A program for causing a computer to function as the image processing apparatus according to any one of claims 1 to 15. An image processing method for generating an output image by reducing noise with respect to an input image, wherein a patch of interest having a first number of pixels is set in the input image, and similar to the patch of interest in the input image. Detecting a plurality of similar patches, and based on the plurality of similar patches, a plurality of target patches including at least one pixel included in the similar patches and having a second number of pixels different from the first number of pixels An image processing method characterized by setting a noise reduction process to the target patch based on the plurality of target patches and synthesizing the target patch whose noise has been reduced; .   The image processing method according to claim 18, wherein the second number of pixels is smaller than the first number of pixels.   The image processing method according to claim 18, wherein a partial region of each of the plurality of similar patches is set as the plurality of target patches.

Images