An Edge-Aware Texture Filtering Method Combining Superpixels

Abstract

The present invention provides an edge-aware texture filtering method combining superpixels, which includes the following steps: Step S1, performing superpixel segmentation on the input image by using the SCAC algorithm to obtain superpixel regions and performing mean filtering with a window size to obtain a mean image; Step S2, calculating the mRTV value for each pixel in the input image, finding the local window corresponding to the minimum mRTV value, and calculating the overlapping part of the local window and the superpixel region to obtain a local filtering region; Step S3, calculating the SPRTV for each pixel in the local filtering region to generate an edge-aware guidance map, calculating the weight for each pixel, and constructing a texture filtering guidance map G' by using the mean image, the edge-aware guidance map, and the weight; Step S4, using G' as a guidance map to perform joint bilateral filtering with the input image to obtain a filtered image; Step S5, proposing evaluation indexes for comparing the similarities and differences between the reference filtered image and the input image and the characteristics of the filtered image itself. This application improves the filtering quality of the existing texture filtering method by organically combining pixel information and superpixel information.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and particularly relates to an edge-aware texture filtering method combining superpixels. Background Art

[0002] Image smoothing is one of the most important preprocessing techniques in the fields of computer vision and graphics. Its purpose is to remove unimportant details such as noise and texture while retaining the significant edge structures of the image. Because it involves the recognition of edge structures and texture details related to image semantics and human perception, texture filtering is a very challenging research problem. Texture filtering has a very wide range of applications, such as structure information extraction, tone mapping, visual abstraction, detail enhancement, and seam carving, and has gradually become a research hotspot in the field of computer vision.

[0003] Traditional edge-preserving smoothing methods are mainly achieved through differences in pixel color or brightness, etc. Therefore, these methods are not very effective in removing high-contrast texture details in images. The global optimization method based on relative total variation proposed by Xu et al. can effectively improve the quality of image smoothing, but this method will have the phenomenon of over-smoothing. Another solution is spatial filtering, which achieves edge-preserving effects by effectively using local windows (local filtering) or similar pixels in the global image (non-local filtering). Compared with local filtering methods, non-local filtering methods further improve the texture smoothing quality by fully considering global similar pixels. However, most of these methods use regular local windows to measure texture details, so the phenomenon of blurring is inevitably present.

[0004] In 2014, Cho et al. proposed bilateral texture filtering. By analyzing texture features with a local regular window and using the idea of window offset to obtain texture information highlighting structural edges from the most representative local window, it can smooth texture details while better retaining structural edges. On this basis, in 2018, Xu et al. proposed edge-aware bilateral texture filtering. By constructing an edge-aware window, each window is made to be as much as possible within a texture region, and the linear features of structural edges are considered in texture measurement. In addition, this method uses a texture filter with a large-scale window and a narrow and long edge-aware small window structure to measure and filter out large-scale textures while retaining fine structures. However, the inventors of the present application have found through research that in essence, the above two methods are both improvements based on regular local filtering windows, and they cannot effectively avoid the blurring phenomenon for fine structural edges, and the edge-preserving ability is not good. Summary of the Invention

[0005] In view of the technical problem that existing texture filtering methods are all improved based on regular local filtering windows and cannot effectively avoid the blurring phenomenon for small structural edges, and the edge retention ability is poor, the present invention provides an edge-aware texture filtering method combining superpixels, which combines pixel information and superpixel information organically to improve the filtering quality of existing texture filtering methods and has better edge retention ability.

[0006] To solve the above technical problems, the present invention adopts the following technical solutions:

[0007] An edge-aware texture filtering method combining superpixels includes the following steps:

[0008] S1. Use the SCAC algorithm to perform superpixel segmentation on the input image I to obtain superpixel regions, and perform mean filtering on the input image I with a window size of r×r to obtain a mean image B;

[0009] S2. Calculate the mRTV value for each pixel p in the input image I, and then find the local window Ω corresponding to the smallest mRTV value p , and calculate the overlapping part of the local window Ω p with the superpixel region to obtain an edge-aware local filtering region SP p ;

[0010] S3. Calculate the SPRTV of each pixel p within the local filtering region SP p to generate an edge-aware guidance map G p , and calculate the weight α for each pixel p, and construct the final texture filtering guidance map G' using the mean image B in step S1, the edge-aware guidance map G p generated in this step, and the weight α;

[0011] S4. Use G' as the guidance map to perform joint bilateral filtering with the input image I to obtain a filtered image J;

[0012] S5. For the filtering result of the texture image, propose evaluation indexes for comparing the similarities and differences between the reference filtered image and the input image and the characteristics of the filtered image itself, so as to better reflect the quality of the image filtering output.

[0013] Further, step S2 specifically includes:

[0014] S21. For each pixel p in the input image I, calculate the mRTV value using the following formula:

[0015]

[0016] where, Δ(Ω p ) represents the chromaticity range within the local window Ω p ; denotes the gradient of the input image I at pixel point j, and ε denotes a very small positive number;

[0017] S22. According to the mRTV values within the r×r neighborhood of each pixel p calculated in step S21, find the local window Ω corresponding to the minimum mRTV value p ;

[0018] S23. Combine the superpixel regions segmented in step S1 and the local window Ω found in step S22 p , and calculate the overlapping part of these two regions for each pixel p to obtain the edge-aware local filtering region SP p .

[0019] Furthermore, step S3 specifically includes:

[0020] S31. Calculate the SPRTV of each pixel p within the local filtering region SP p using the following formula:

[0021]

[0022] where, denotes the gradient of the input image I at pixel point j, and r is the scale size of the offset window;

[0023] S32. Calculate the weight α for each pixel p using the following formula:

[0024]

[0025] where, SPRTV(Ω p ) represents the SPRTV value of the local window Ω p , and SPRTV(SP p ) represents the SPRTV value of the local filtering region SP p , and σ represents the conversion weight from the edge to the texture region;

[0026] S33. Calculate the mean value of the pixels in the edge-aware local filtering region SP p to generate a new guidance map G p , and construct the final texture filtering guidance map G' using the mean image B in step S1 and the weight α calculated in step S32 through the formula G p ' = α p G p + (1 - α p )B p .

[0027] Furthermore, step S5 specifically includes:

[0028] S51. For multiple images in the publicly available existing texture image dataset, perform edge detection using a very small threshold through the Canny operator to obtain all edge information of the structure and details.

[0029] S52. Subtract the structural edges provided in the existing texture image dataset from the edge information obtained in step S51 to separately obtain the image structural edge and the detail information edge.

[0030] S53. Calculate the gradients in the x and y directions at each pixel point for the input image I and the filtered image J respectively, measure the similarities and differences in the gradients of the filtered image and the input image at the structural edge and the detail edge respectively, and finally integrate them into a unified evaluation index E F :

[0031]

[0032]

[0033]

[0034] where S and T respectively represent the structural edge and the detail information pixel sets, I x (i) and I y (i) are the gradients of the original input image pixel point i, J x (j) and J y (j) are the gradients of the filtered image pixel point j, J x (i) and J y (i) are the gradients of the filtered image pixel point i, and the filtered image pixel point j is the neighborhood of i, and N(i) represents the 3×3 neighborhood centered on the pixel point i.

[0035] Compared with the prior art, the edge-aware texture filtering method based on joint superpixels provided by the present invention has the following advantages:

[0036] 1. The present invention uses the SCAC algorithm to perform superpixel region segmentation on the input image, thereby constructing a locally filtered region with edge awareness, and improving the problem of edge blurring caused by most existing texture filtering methods using regular rectangular windows. For each pixel, using the characteristics of superpixels, that is, by means of feature measurement, the same or similar local pixels are aggregated into a superpixel block, and combined with window offset, a locally filtered region with stronger edge awareness is constructed. For the small structural edges in the image, the pixels included in the edge-aware window proposed by the present invention are more representative, so that the structural edges of the image can be better retained.

[0037] 2. In generating the bilateral texture filtering guidance map, the present invention improves the calculation method of the RTV value and improves texture measurement based on the edge-aware window. Then, the present invention proposes a new weight calculation method to further improve the quality of the guidance map, thereby enhancing the ability to retain structural edges.

[0038] 3. When evaluating the quality of image texture filtering, the present invention obtains an evaluation index that can better reflect the image filtering quality by comparing the similarities and differences between the filtered image and the original input image as well as the characteristics of the filtered image itself. Brief Description of the Drawings

[0039] Figure 1 is a schematic flowchart of the edge-aware texture filtering method for joint superpixels provided by an embodiment of the present invention.

[0040] Figure 2 is a comparison diagram of the results of superpixel segmentation by the SLIC algorithm and the SCAC algorithm provided by an embodiment of the present invention.

[0041] Figure 3 is a filtering result diagram of the method (h) of the present invention and classical filtering methods: the filtering method of relative total variation (b), bilateral texture filtering (c), iterative guided filtering (d), tree filtering (e), Gaussian correlation texture filtering (f), side window filtering (g) provided by an embodiment of the present invention.

[0042] Figure 4 is a filtering result diagram of the method (e) of the present invention and scale-aware texture filtering methods: scale-adaptive texture filtering (b), fast scale-adaptive bilateral texture filtering (c), scale-aware texture filtering (d) provided by an embodiment of the present invention.

[0043] Figure 5 is a filtering result diagram of the woman image obtained by respectively using the filtering method of relative total variation (a), bilateral texture filtering (b), iterative guided filtering (c), tree filtering (d), Gaussian correlation texture filtering (e), side window filtering (f), scale-adaptive texture filtering (g), fast scale-adaptive bilateral texture filtering (h), scale-aware texture filtering (i) and the method (j) proposed by the present invention provided by an embodiment of the present invention.

[0044] Figure 6 is a filtering result diagram of the texture image obtained by respectively using the filtering method of relative total variation (c), bilateral texture filtering (d), iterative guided filtering (e), tree filtering (f), Gaussian correlation texture filtering (g), side window filtering (h), scale-adaptive texture filtering (i), fast scale-adaptive bilateral texture filtering (j), scale-aware texture filtering (k) and the method (b) proposed by the present invention provided by an embodiment of the present invention.

[0045] Figure 7 It is the texture information obtained by using the Canny operator and the provided structural edges in the embodiments of the present invention.

[0046] Figure 8 It is the mean comparison graph obtained by using the above ten filtering methods on 200 images with the proposed evaluation index in the embodiments of the present invention. Specific embodiments

[0047] In order to make the technical means, creative features, achieved purposes and effects of the present invention easy to understand, the present invention will be further described below with reference to specific illustrations.

[0048] Please refer to Figure 1 As shown, the present invention provides an edge-aware texture filtering method combining superpixels, including the following steps:

[0049] S1. Use the SCAC algorithm to perform superpixel segmentation on the input image I to obtain superpixel regions, and perform mean filtering on the input image I with a window size of r×r to obtain a mean image B. Specifically, a superpixel refers to an irregular pixel block with certain visual significance composed of local pixels with similar texture, color, brightness, etc. The superpixel technology groups pixels using the similarity between pixels, so that similar pixels are aggregated into a superset. The SLIC (Simple Linear Iterative Clustering) algorithm can generate compact and approximately uniform superpixels, and has high performance in terms of operation speed, object contour preservation, superpixel shape, etc., meeting the segmentation effect expected by human vision. However, the SLIC algorithm only divides superpixels based on the color information of the image, and performs iterative segmentation through a clustering algorithm that continuously updates the seed points, and is not well applicable to the fine edges in the image, especially for images with complex textures. In 2021, Yuan et al. proposed a content-adaptive superpixel segmentation method (SCAC), which classifies the content in the image through gradient, color, and texture information, performs adaptive superpixel segmentation on different information respectively, and finally aggregates the segmentation results of each type to obtain the final content-adaptive superpixel segmentation image. Compared with the SLIC algorithm, the SCAC algorithm has better superpixel segmentation results for texture images, especially at the texture regions and the edges of object structures. As Figure 2 shown, the white area is the overlapping part of the offset window and the superpixel region, that is, the edge-aware local filtering region SP proposed by the present invention p, where (a) is the input image, (b) is the segmentation result of the SLIC algorithm, and (c) is the segmentation result of the SCAC algorithm. It can be seen that the segmentation using the SCAC algorithm is more accurate. Therefore, the present invention finally uses the SCAC algorithm to segment the input image I to obtain the superpixel region, and performs mean filtering on the input image I with a window size of r×r to obtain the mean image B.

[0050] S2. Calculate the mRTV value for each pixel p in the input image I, and then find the local window Ω corresponding to the minimum mRTV value p , calculate the local window Ω p and the overlapping part of the superpixel region to obtain the edge-aware local filtering region SP p . Considering the strategy of the bilateral texture filtering local filtering window, the edge-aware local filtering region proposed by the present invention is generated by combining irregular superpixels and window offsets. The window offset selects a regular local window for the pixel that is least likely to contain the structural edge, and then calculates the overlapping part with the superpixel to obtain an irregular local filtering region that does not contain the structural edge, avoiding the blurring phenomenon caused by the filtering window crossing the structural edge. Specifically, step S2 includes the following steps:

[0051] S21. For each pixel p in the input image I, calculate the mRTV value using the following formula:

[0052]

[0053] where, Δ(Ω p ) represents the chromaticity range within the local window Ω p , represents the gradient of the input image I at the pixel point j, and ε represents a very small positive number.

[0054] S22. According to the mRTV values within the r×r neighborhood of each pixel p calculated in step S21, find the local window Ω corresponding to the minimum mRTV value p . In theory, a smaller mRTV value indicates that the pixel points contained in this region belong to the same or similar texture structures. Therefore, select the local window Ω corresponding to the minimum mRTV value p , so that the regular rectangular window does not contain pixels with large differences as much as possible, thereby weakening the influence brought by the different pixels.

[0055] S23. Combine the superpixel region segmented in step S1 and the local window Ω found in step S22 p , calculate the overlapping part of these two regions for each pixel p to obtain the edge-aware local filtering region SP p , and this region shows irregularity around the structural edge, showing the characteristics of edge awareness.

[0056] S3. Calculate the SPRTV of each pixel p within the local filtering region SP p to generate an edge-aware guidance map G, and calculate the weight α for each pixel p. Use the mean image B from step S1 and the edge-aware guidance map G generated in this step p along with the weight α to construct the final texture filtering guidance map G'. Specifically, given the input image I, combined with the intermediate guidance map G, the texture filtering result image is obtained through joint bilateral filtering: p where k

[0057]

[0058] is the normalization term, Ω p is the local window centered on pixel p, and the output J p is the weighted average of I p within the neighborhood Ω q of pixel p. The spatial position kernel f and the color range kernel g are typical Gaussian functions. The color range kernel g is inversely proportional to the feature difference between two pixels p and q of the guidance map G. Therefore, the quality of the guidance map has a great impact on the filtering result. An ideal guidance map should have the following two characteristics: First, the pixel gradient in the texture area should be small enough, which is beneficial to texture removal; second, the gradient at the structural edge should be consistent with the original image, showing discontinuity and retaining the structural features of the original image. p Since a regular rectangular filtering window is used, there are still large errors for pixels on fine structural edges. In addition, for noise with a very high chromaticity value, since it may have a high RTV (relative total variation) value, it may also be recognized as edge noise. To improve the above phenomenon, this step S3 specifically includes the following steps:

[0059] S31. Calculate the SPRTV of each pixel p within the local filtering region SP

[0060] by the following formula: p where SP

[0061]

[0062] is the fusion window offset corresponding to pixel p and the edge-aware local filtering region obtained by superpixel segmentation, p represents the gradient of the input image I at pixel point j, and r is the scale size of the offset window.

[0063] S32. Based on the above step S31, strengthen the calculation of the weight α to improve the quality of the guidance map; the improved formula for calculating the weight α is as follows, that is, the weight α for each pixel p is calculated by the following formula:

[0064] ​

[0065] Among them, Ω p represents a local window centered on pixel p, and SPRTV(Ω p ) represents the SPRTV value of the local window Ω p . SPRTV(SP p ) represents the SPRTV value of the local filtering region SP p . σ represents the conversion weight from the edge to the texture region.

[0066] For the pixel points on the image edge, the value of SPRTV(SP p ) is smaller than that of SPRTV(Ω p ) because the regular rectangular offset window Ω p contains structural edges, while the pixels in the edge-aware local filtering region SP p belong to the same texture region, which will make the improved weight α value larger, so the structural edges of the image can be maintained.

[0067] For the pixel points on the texture edge in the image, the situation is similar to that of the pixel points at the structural edge, and its weight α value will also be larger. However, since the surrounding pixels also belong to the texture region, and as the number of iterations increases, the gradient of this pixel will decrease rapidly, so the effect of removing the texture edge can be achieved.

[0068] For the pixel points in the flat region of the image, the values of SPRTV(SP p ) and SPRTV(Ω p ) are very close. Through the above formula, the weight α value will be smaller, so the texture information of the image can be efficiently removed.

[0069] S33. Calculate the mean value of the pixels in the edge-aware local filtering region SP p to generate a new guidance map G p , and use the mean image B in step S1 and the weight α calculated in step S32 to construct the final texture filtering guidance map G' through the formula G p ' = α p G p + (1 - α p )B p .

[0070] S4. Use G' as the guidance map to perform joint bilateral filtering with the input image I to obtain the filtered image J.

[0071] S5. For the filtering result of the texture image, propose evaluation indexes for comparing the similarities and differences between the reference filtering image and the input image and the characteristics of the filtering image itself, so as to better reflect the quality of the image filtering output. Specifically, step S5 includes the following steps:

[0072] S51. For multiple images in the publicly available existing texture image dataset (such as 200 images in the texture image dataset provided by Xu Li in 2012), perform edge detection using a very small threshold through the Canny operator to obtain all edge information of the structure and details.

[0073] S52. Subtract the structural edges provided in the existing texture image dataset from the edge information obtained in step S51 to separately obtain the image structural edge and the detail information edge. Please refer to Figure 7 As shown, (a) is the input image, (b) is all the edge information obtained by performing edge extraction on the input image using the Canny operator, (c) is the structural edge provided in the dataset, and (d) is the texture information edge obtained by subtraction.

[0074] To quantitatively measure the filtering effect of the method proposed in the present invention and other existing filtering methods for comparison, the currently used evaluation metrics were studied. However, the evaluation metrics used in a large number of studies are the peak signal-to-noise ratio (PSNR) and the structural similarity (SSIM), but these two evaluation metrics are not applicable to texture filtering methods. Considering the difference between the filtered image and the input image and combining the characteristics of the filtered output image itself. The visual understanding is that at the texture information, the filtered output image should have a large gradient difference from the input image because the texture information should be smoothed; at the edge structure, the filtered output image should be consistent with the input image, so the gradient difference is small because the edge structure needs to be retained. Furthermore, the characteristics of the filtered output image itself should also be considered. The gradient at its texture edge not only has a large difference from the output image, but its own gradient should be small; similarly, the gradient at the structural edge not only has a small difference from the output image, but its own gradient should be large. Therefore, the present invention proposes an evaluation metric E applicable to texture filtering methods F .

[0075] S53. Calculate the gradients in the x and y directions at each pixel point for the input image I and the filtered image J respectively to measure the similarities and differences in gradients at the structural edge and the detail edge between the filtered image and the input image, and finally integrate them into a unified evaluation metric E F :

[0076]

[0077]

[0078]

[0079] where S and T respectively represent the pixel sets of the structural edge and the detail information, I x (i) and I y(i) is the gradient of the original input image pixel i, J x (j) and J y (j) are the gradients of the filtered image pixel j, J x (i) and J y (i) are the gradients of the filtered image pixel i, and the filtered image pixel j is the neighborhood of i, and N(i) represents the 3×3 neighborhood centered on pixel i.

[0080] From the above formula for calculating E S , it can be known that if the structural edges can be well preserved, the value of E S is small; otherwise, the value of E S will be large. From the above formula for calculating E T , it can be known that if the texture edges can be effectively removed, the value of E S is large; otherwise, the value of E S will be small. Therefore, in the above formula for calculating E F , the larger the value of E F , the better the filtering performance of the algorithm. Please refer to Figure 8 shown in the figure, which is a comparison chart of using the evaluation index proposed by the present invention for different filtering methods on a dataset of 200 collected images. It can be seen from it that the method proposed by the present invention has a better performance in this index.

[0081] Please refer to Figure 3 shown in the figure, which is the filtering result chart of using the filtering method of relative total variation (b), bilateral texture filtering (c), iterative guided filtering (d), tree filtering (e), Gaussian correlation texture filtering (f), side window filtering (g) and the method proposed by the present invention (h). It can be seen from the enlarged view that the method proposed by the present invention can well remove the texture information in the image while retaining the structural edges of the image.

[0082] Please refer to Figure 4 shown in the figure, which is the filtering result chart of using scale-adaptive texture filtering (b), fast scale-adaptive bilateral texture filtering (c), scale-aware texture filtering (d) and the method of the present invention (e). The scale-adaptive texture filtering and fast scale-adaptive bilateral texture filtering methods both adaptively calculate the optimal filtering scale for each pixel through directional gradient information. Since they still use regular rectangular windows and only set different scale radii for pixels in different regions, they cannot effectively retain small edge structures. The scale-aware texture filtering method weakens the gradients of small-scale textures with high contrast through scale-aware filtering, combined with L0H -1The variational model preserves the structure while implementing the filtering operation. However, the L0 gradient minimization model highly depends on the gradient information of the image and cannot well perceive the fine structural edges, resulting in blurring. The method proposed in the present invention can well alleviate the above problems, achieve the purpose of preserving the structural edges while removing the texture, and obtain a good filtering effect.

[0083] Please refer to Figure 5 as shown, the filtering result graphs obtained by respectively using the relative total variation filtering method (a), bilateral texture filtering (b), iterative guided filtering (c), tree filtering (d), Gaussian correlation texture filtering (e), side window filtering (f), scale-adaptive texture filtering (g), fast scale-adaptive bilateral texture filtering (h), scale-aware texture filtering (i) and the method proposed in the present invention (j) for the woman image. Compared with other methods, the method proposed in the present invention can better retain the structural edges in the part of the head ornaments; in the clothes pattern area at the lower left corner, the method proposed in the present invention can retain the structural edges while smoothing the details.

[0084] Please refer to Figure 6 as shown, the filtering result graphs obtained by respectively using the relative total variation filtering method (c), bilateral texture filtering (d), iterative guided filtering (e), tree filtering (f), Gaussian correlation texture filtering (g), side window filtering (h), scale-adaptive texture filtering (i), fast scale-adaptive bilateral texture filtering (j), scale-aware texture filtering (k) and the method proposed in the present invention (b) for the texture image. It can be clearly seen that the method proposed in the present invention can well remove the texture information and retain the structural edges of the image, and has a better visual effect compared with other filtering methods.

[0085] Compared with the prior art, the edge-aware texture filtering method combining superpixels provided by the present invention has the following advantages:

[0086] 1. The present invention uses the SCAC algorithm to segment the input image into superpixel regions, thereby constructing a locally filtering region with edge awareness, and improving the problem of edge blurring caused by most of the existing texture filtering methods using regular rectangular windows. For each pixel, by using the characteristics of superpixels, that is, by the feature measurement method, the same or similar local pixels are aggregated into a superpixel block, and combined with the window offset, a more edge-aware local filtering region is constructed. For the fine structural edges in the image, the pixels included in the edge-aware window proposed in the present invention are more representative, so as to better retain the structural edges of the image.

[0087] 2. In generating the bilateral texture filtering guidance map, the present invention improves the calculation method of the RTV value and improves the texture measurement based on the edge-aware window. Then, the present invention proposes a new weight calculation method to further improve the quality of the guidance map, thereby enhancing the ability to retain structural edges.

[0088] 3. When evaluating the quality of image texture filtering, the present invention obtains an evaluation index that can better reflect the image filtering quality by comparing the similarities and differences between the filtered image and the original input image and the characteristics of the filtered image itself.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the purpose and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. An edge-aware texture filtering method combining superpixels, characterized in that, It includes the following steps: S1. Use the SCAC algorithm to perform superpixel segmentation on the input image I to obtain superpixel regions, and perform mean filtering on the input image I with a window size of r×r to obtain a mean image B; S2. Calculate the mRTV value for each pixel p in the input image I, and then find the local window Ω corresponding to the minimum mRTV value p , calculate the local window Ω p The overlapping part of the local window Ω p ; S3. Calculate the SPRTV of each pixel p in the local filtering region SP p to generate an edge-aware guidance map G p , and calculate the weight α for each pixel p. Use the mean image B in step S1, the edge-aware guidance map G generated in this step p and the weight α to construct the final texture filtering guidance map G'. S4. Use G' as a guidance map to perform joint bilateral filtering with the input image I to obtain a filtered image J; S5. For the filtering results of the texture image, propose evaluation indexes for comparing the similarities and differences between the reference filtered image and the input image and the characteristics of the filtered image itself, which are used to better reflect the quality of the image filtering output; Among them, the specific steps of step S3 include: S31. Calculate the SPRTV of each pixel p in the local filtering region SP p using the following formula: Among them, represents the gradient of the input image I at pixel point j, and r is the scale size of the offset window; S32. Calculate the weight α for each pixel p using the following formula: Among them, SPRTV(Ω p ) represents the SPRTV value of the local window Ω p , and SPRTV(SP p ) represents the SPRTV value of the local filtering region SP p . σ represents the conversion weight from the edge to the texture region; S33. Calculate the mean value of the pixels in the local filtering region SP with edge perception to generate a new guidance map G p and construct the final texture filtering guidance map G' by using the mean value image B in step S1 and the weight α calculated in step S32 through the formula G p ' = α p G p +(1 - α p )B p . p ​ 2. The edge-aware texture filtering method for joint superpixels according to claim 1, characterized in that, The specific steps of step S2 include: S21. For each pixel p in the input image I, calculate the mRTV value using the following formula: where, Δ(Ω p ) represents the chromaticity range within the local window Ω p , represents the gradient of the input image I at pixel point j, and ε represents a very small positive number; S22. Based on the mRTV values within the r×r neighborhood of each pixel p calculated in step S21, find the local window Ω corresponding to the minimum mRTV value p ; S23. Combine the superpixel regions obtained in step S1 and the local window Ω found in step S22 p , calculate the overlapping part of these two regions for each pixel p to obtain a locally filtered region SP with edge awareness p .

3. The edge-aware texture filtering method for jointly superpixels according to claim 1, characterized in that The specific steps of step S5 include: S51. For multiple images in the publicly available existing texture image dataset, perform edge detection using an extremely small threshold through the Canny operator to obtain all edge information of the structure and details; S52. Subtract the structural edges provided in the existing texture image dataset from the edge information obtained in step S51 to separate the image structural edge and the detail information edge; S53. Calculate the gradients in the x and y directions at each pixel point of the input image I and the filtered image J respectively, measure the similarities and differences in gradients between the filtered image and the input image at the structural edges and detail edges respectively, and finally integrate them into a unified evaluation index E F : Among them, S and T respectively represent the structural edge and the detail information pixel set, and I x (i) and I y (i) are the gradients of the original input image pixel point i, and J x (j) and J y (j) are the gradients of the filtered image pixel point j, and J x (i) and J y (i) are the gradients of the filtered image pixel point i, and the filtered image pixel point j is the neighborhood of i, and N(i) represents the 3×3 neighborhood centered on the pixel point i.

Images