Self-adaptive grid division image splicing method based on image texture information distribution

Abstract

The invention discloses a self-adaptive grid division image splicing method based on image texture information distribution. The method specifically comprises the steps that feature points of images are extracted and matched; calculating a global homography matrix by using the feature point pairs, applying the global homography matrix to the target image, and generating a coarse registration image; calculating a gray level co-occurrence matrix and a contrast ratio, obtaining a texture information richness degree, and segmenting the image into non-uniform grids based on the texture information richness degree; calculating a grid vertex homography matrix by using a feature point matching pair, and applying different weights; and generating a deformed target image by using moving homography transformation, and processing the reference image and the transformed target image by using a feather fusion method to generate a fused image. According to the method, the distribution condition of the image texture information is fully considered, and the texture information is used for guiding grid division, so that the alignment effect of the image overlapping region is better, and meanwhile, the robustness is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image splicing methods, and in particular relates to an image splicing method based on adaptive grid division and image texture information distribution. Background Art

[0002] Image stitching technology, a key research area in computer vision, demonstrates significant technical value and promising development prospects in multiple application areas. Image stitching involves aligning multiple overlapping images to generate a single wide-field-of-view image, simplifying the processing of advanced visual tasks such as object recognition and anomaly detection. However, most existing image stitching methods use a uniform grid to partition images. This method assumes that the texture information distribution of an image is uniform and that the texture richness is within an appropriate range. Grid division is a crucial step in image stitching, and the resulting grid division directly impacts the stitching quality. The number of grid cells is generally adjusted based on the size of the stitched images, stitching speed, and stitching accuracy requirements. However, the texture information distribution varies significantly between images. Using a uniform grid to partition images ignores the impact of image texture on the stitching quality. When the texture information distribution of images varies significantly, using a uniform grid to partition images is affected by the number of grid cells. Excessive grid density significantly increases computational redundancy and reduces stitching speed. Setting a dense grid in smooth textured areas has little effect on the stitching quality. However, low grid density means that using the same homography for projection transformation within a single grid in texture-rich areas can make it difficult to align texture features, affecting the stitching effect. Furthermore, manually calculating the number of grids for each set of images to be stitched increases the image stitching workload, making it difficult to apply to stitching multiple sets of images. Summary of the Invention

[0003] The purpose of the present invention is to provide an adaptive grid division image stitching method based on image texture information distribution, which solves the problems of high computational complexity and artifacts in the stitched images in existing image stitching methods.

[0004] The technical solution adopted by the present invention is an adaptive grid division image splicing method based on image texture information distribution, which is specifically implemented according to the following steps:

[0005] Step 1: Extract and match the feature points of the reference image and the target image;

[0006] Step 2: Calculate the global homography matrix by matching feature points, pre-register the two images, and generate a coarse registration image;

[0007] Step 3: Calculate the gray level co-occurrence matrix of the coarse registration image, use contrast to represent the image texture information, and obtain the texture information of different areas of the image;

[0008] Step 4: After calculating the contrast in different grid areas, compare it with the threshold, and recursively divide it to obtain the grid vertex coordinates, grid width, and grid height;

[0009] Step 5: Calculate the transformation matrix of different mesh vertices based on the feature point matching pairs and the feature point weights, and deform the mesh vertices according to the transformation matrix;

[0010] Step 6: Use the transformation matrix of the mesh vertices to guide the transformation of the pixels within the mesh to generate the deformed target image;

[0011] Step 7: Fuse the deformed target image with the reference image and use the feathering fusion method to eliminate the image boundaries.

[0012] The present invention is also characterized in that:

[0013] In step 1, specifically:

[0014] For the target image I r Deform it to the reference image I l The SIFT algorithm is used to extract the feature points of the two images and generate feature descriptors. The reference image I is l and target image I r The feature points in the image are matched, and then the RANSAC method is used to eliminate the mismatched pairs.

[0015] In step 2, specifically:

[0016] Using at least four sets of feature point matching pairs and any three sets of feature points are not collinear, the global homography transformation matrix H is calculated. The specific formula is shown in formula (1):

[0017]

[0018] Where: x′ i Denotes the reference image I l Feature point horizontal coordinate, y′ i Denotes the reference image I l Feature point vertical coordinate, x i Represents the target image I r Feature point horizontal coordinate, y i Represents the target image I r Feature point vertical coordinate, h 00 ,h 01 ,…h 22 are the 9 parameters of the homography transformation matrix H;

[0019] Expanding formula (1) yields formula (2):

[0020]

[0021] The global homography transformation matrix is ​​calculated by solving equation (2) using the least squares method. The target image is deformed using the global homography transformation matrix and fused with the reference image to obtain the coarse registration image I g .

[0022] In step 3, the details are as follows:

[0023] Step 3.1, calculate the coarse registration image I g Gray level co-occurrence matrix, as shown in formula (3); set the contrast threshold Th, grid width upper limit W max and the grid width lower limit W min ;

[0024]

[0025] Where: P(i,j) is the element in the gray-level co-occurrence matrix, i and j refer to the pixels with gray-level values ​​i and j respectively; C(i,j|d,q) is the number of occurrences of pixel pairs with gray-level values ​​i and j, d represents the spatial distance, θ represents the direction; N represents the number of gray-level levels;

[0026] Step 3.2: The coarse registration image is set according to the upper limit of the grid width W. max Perform pre-division, and then calculate the contrast of the coarse registration image in different grid areas through the gray level co-occurrence matrix, as shown in formula (4);

[0027]

[0028] Among them: Contrast represents the contrast.

[0029] In step 4, specifically:

[0030] After calculating the contrast in different grid areas, compare it with the contrast threshold Th. If it is greater than the threshold, the current grid is divided into 4 sub-areas until the contrast in the grid is less than the threshold or the grid width is less than W. min ;Store the current grid point information, including grid vertex coordinates, grid width and grid height.

[0031] In step 5, specifically:

[0032] Step 5.1, assign different weights to each feature point according to the distance from the grid center point to the feature point;

[0033] The weight calculation formula is shown in formula (5);

[0034]

[0035] Where: x *represents the center point of the grid, x i represents the characteristic point, σ is the bandwidth parameter;

[0036] Weight Matrix As shown in formula (6);

[0037]

[0038] Step 5.2, use the feature matching point pairs to calculate the homography transformation matrix of each grid center point. The homography matrix calculation is shown in Equations (7) and (8);

[0039]

[0040] Where: is the coordinate of the center point of the grid after transformation, H * is the local homography transformation matrix, x * ′ is the grid center coordinate of the coarse registration image, h * Denotes the local homography matrix H * The parameterized form of Represents the Gaussian weight of the i-th feature point to the current local area, a i It is a constraint matrix constructed based on the i-th feature point matching pair, converting the homography matrix into a linear constraint.

[0041] In step 6, the homography matrices for the different grids obtained in step 5 are used to calculate the pixel position of each pixel on the canvas after the homography transformation. If the pixel is within the target image, the pixel is copied from the target image to the canvas. If the mapped coordinates are not integers, they are rounded before processing.

[0042] In step 7, the feathering fusion algorithm is used to eliminate the image boundary; weights are assigned to pixels according to the distance from the pixel to the boundary. The feathering fusion formula is shown in formula (9);

[0043] I(x,y)=ω1(x,y)×I1(x,y)+ω2(x,y)×I2(x,y) (9)

[0044] Where I(x,y) represents the grayscale value of the pixel in the boundary area of ​​the image, I1(x,y) and I2(x,y) are the grayscale values ​​of the edge pixels of the reference image and the target image on both sides of the boundary, ω1(x,y) and ω2(x,y) are weight functions and satisfy ω1(x,y)+ω2(x,y)=1;

[0045] The weight is calculated based on the distance from the pixel to the boundary, as shown in formula (10);

[0046]

[0047] For the reference image I1, d1(x,y) is the closest distance from the pixel point (x,y) to the boundary of I1, and d2(x,y) is the closest distance from the pixel point (x,y) to the boundary of I2.

[0048] The beneficial effects of the present invention are:

[0049] Compared with the traditional dense grid method, the method of the present invention can reduce the amount of calculation in the image stitching process and improve the image stitching speed; in addition, by using texture-guided grid division combined with local homography transformation, it can significantly reduce artifacts in large parallax scenes and improve the visual effect of the stitched image. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is the overall architecture diagram of the adaptive grid division and image splicing method based on image texture information distribution of the present invention;

[0051] Figure 2 Schematic diagram of the gray-level co-occurrence matrix of the present invention;

[0052] Figure 3 is a flow chart of adaptive meshing of the present invention;

[0053] Figure 4 It is a schematic diagram of the adaptive grid division of the present invention;

[0054] Figure 5a is the reference image in Example 3;

[0055] Figure 5b is the target image in Example 3;

[0056] Figure 5c is a schematic diagram of image feature point matching pairs in Example 3;

[0057] Figure 5d is the coarse registration image in Example 3;

[0058] Figure 5e This is the adaptive meshing result diagram in Example 3;

[0059] Figure 5f is the target image after deformation in Example 3;

[0060] Figure 5g is the stitched image in Example 3;

[0061] Figure 6a is the reference image in Example 3;

[0062] Figure 6b is the target image in Example 3;

[0063] Figure 6cis a schematic diagram of image feature point matching pairs in Example 3;

[0064] Figure 6d is the coarse registration image in Example 3;

[0065] Figure 6e This is the adaptive meshing result diagram in Example 3;

[0066] Figure 6f is the target image after deformation in Example 3;

[0067] Figure 6g is the stitched image in Example 3;

[0068] Figure 7a Root mean square error comparison diagram of the method of the present invention and other methods;

[0069] Figure 7b This is a comparison chart of the peak signal-to-noise ratio between the method of the present invention and other methods;

[0070] Figure 7c It is a structural similarity comparison diagram between the method of the present invention and other methods. DETAILED DESCRIPTION

[0071] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0072] Example 1

[0073] The present invention is based on the adaptive grid division image splicing method of image texture information distribution, such as Figure 1 As shown, please follow the steps below:

[0074] Step 1: Image feature point extraction and matching: First, use SIFT (Scale Invariant Feature Transform) to extract image feature points and generate feature descriptors. Use the feature descriptors to match the images to be stitched to generate several sets of feature point matching pairs. Second, use the RANSAC (Random sample consensus) method to remove mismatched pairs to facilitate the subsequent calculation of the transformation relationship between images.

[0075] Specifically: Input the image pair to be stitched, which are the reference image I l and target image I r In the image stitching algorithm, the target image needs to be deformed to transform it to the perspective of the reference image. The SIFT algorithm is used to extract the feature points of the two images and generate feature descriptors. l and I rThen the RANSAC method is used to eliminate mismatched pairs to improve the accuracy of the subsequent global homography transformation matrix.

[0076] Step 2: Image pre-registration: Calculate the global homography matrix by matching multiple feature point pairs; then use the global homography matrix to pre-register the two images to generate a coarsely registered image, specifically:

[0077] The global homography transformation matrix H is calculated using feature point matching pairs (at least four groups, and any three groups of feature points are not collinear). The specific formula is shown in formula (1):

[0078]

[0079] Where: x′ i Denotes the reference image I l Feature point horizontal coordinate, y′ i Denotes the reference image I l Feature point vertical coordinate; x i Represents the target image I r Feature point horizontal coordinate, y i Represents the target image I r The vertical coordinate of the feature point. h 00 ,h 01 ,…h 22 are the 9 parameters of the homography transformation matrix H, (h 00 ,h 01,10 ,h 11 The four parameters jointly control the linear transformations of the image such as rotation, scaling and shearing, h 02 ,h 12 Represents the translation along the x and y directions respectively, and controls the translation of the image together. 20 ,h 21 Control the perspective effect of the image, h 22 is the homogeneous scale factor. Due to the scale invariance of homogeneous coordinates, h is usually 22 Fixed to 1);

[0080] Expanding formula (1) yields formula (2):

[0081]

[0082] Where: (x i ,y i ) and (x′ i ,y′ i ) is a pair of feature point matching pairs, i∈[1,n], and n is the total number of feature point matching pairs.

[0083] The global homography transformation matrix is ​​calculated by solving equation (2) using the least squares method. The target image is deformed using the global homography transformation matrix and fused with the reference image to obtain the coarse registration image I g ;

[0084] Step 3, obtain the texture information of the pre-registered image: calculate the gray level co-occurrence matrix of the coarsely registered image, use contrast to represent the image texture information, and obtain the texture information of different areas of the image. Figure 3 As shown, the details are as follows:

[0085] Step 3.1, calculate the coarse registration image I g Gray-level co-occurrence matrix, such as Figure 2 As shown, the contrast threshold Th is set, which is usually the global contrast of the coarse registration image and the upper limit of the grid width W. max (usually 0.1 of the image width) and the grid width lower limit W min (Usually 0.01 of the image width).

[0086]

[0087] Where: P(i,j) is the element in the gray-level co-occurrence matrix, i and j refer to the pixels with gray-level values ​​i and j respectively; C(i,j|d,q) is the number of occurrences of pixel pairs with gray-level values ​​i and j, d represents the spatial distance, and θ represents the direction. N represents the number of gray-level levels.

[0088] Step 3.2: The coarse registration image is set according to the upper limit of the grid width W. max After pre-division, the contrast of the coarsely registered images in different grid areas is calculated using the gray-level co-occurrence matrix, as shown in formula (4).

[0089]

[0090] Among them: Contrast represents the contrast;

[0091] Step 4, adaptive grid division: recursively divide the coarse registration image by setting the texture richness threshold and the grid width threshold until the texture information richness of each grid in the coarse registration image is lower than the threshold or the grid width is smaller than the preset range; Figure 4 As shown;

[0092] After calculating the contrast in different grid areas, compare it with the preset threshold Th. If it is greater than the threshold, the current grid is divided into 4 sub-areas until the contrast in the grid is less than the threshold or the grid width is less than W. min .Store the current grid point information, including grid vertex coordinates, grid width and grid height.

[0093] Step 5, mesh vertex deformation: After obtaining the mesh vertex information, calculate the transformation matrix of different mesh vertices based on the feature point matching pairs and feature point weights, and then deform the mesh vertices according to the transformation matrix;

[0094] Step 5.1: To improve the image alignment accuracy, assign different weights to each feature point based on the distance from the grid center point to the feature point. The farther the feature point is from the grid center point, the smaller its weight is.

[0095] The weight calculation formula is shown in formula (5);

[0096]

[0097] Where: x * represents the center point of the grid, x i Represents the feature point, σ is a bandwidth parameter used to control the weight decay rate, generally set to 1.

[0098] Weight Matrix As shown in formula (6);

[0099]

[0100] Step 5.2, use the feature matching point pairs to calculate the homography transformation matrix of each grid center point. The homography matrix calculation is shown in Equations (7) and (8);

[0101]

[0102] Where: is the coordinate of the center point of the grid after transformation, H * is the local homography transformation matrix, x * ′ is the grid center coordinate of the coarse registration image, h * Denotes the local homography matrix H * The parameterized form of Represents the Gaussian weight of the i-th feature point to the current local area, a i It is a constraint matrix constructed based on the i-th feature point matching pair, converting the homography matrix into a linear constraint.

[0103] Step 6: Target image deformation: Use the transformation matrix of the mesh vertices to guide the transformation of the pixels within the mesh to generate the deformed target image;

[0104] Using the homography matrices for the different grids obtained in step 5, calculate the pixel position of each pixel on the canvas after homography transformation. If the pixel is within the target image, copy the pixel from the target image to the canvas. If the mapped coordinates are not integers, round them up before processing.

[0105] Step 7, image fusion: fuse the deformed target image with the reference image, and use the feathering fusion method to eliminate the image boundaries;

[0106] Use the feathering fusion algorithm to eliminate image boundaries. Pixels are assigned weights based on their distance from the boundary. Pixels closer to the boundary have smaller weights, while pixels farther away have larger weights. The feathering fusion formula is shown in Equation (9);

[0107] I(x,y)=ω1(x,y)×I1(x,y)+ω2(x,y)×I2(x,y) (9)

[0108] Where I(x,y) represents the grayscale value of the pixel in the image boundary area, I1(x,y) and I2(x,y) are the grayscale values ​​of the edge pixels of the reference image and the target image on both sides of the boundary, and ω1(x,y) and ω2(x,y) are weight functions that satisfy ω1(x,y)+ω2(x,y)=1. The weight is calculated based on the distance from the pixel to the boundary, as shown in Equation (10).

[0109]

[0110] For the reference image I1, d1(x,y) is the closest distance from the pixel point (x,y) to the boundary of I1, and d2(x,y) is the closest distance from the pixel point (x,y) to the boundary of I2.

[0111] Example 2

[0112] This adaptive grid-based image stitching method, based on the distribution of image texture information, first extracts and matches feature points from the image. Next, a global homography matrix is ​​calculated using feature point pairs and applied to the target image to generate a coarsely registered image. The grayscale co-occurrence matrix and contrast ratio are then calculated to determine the richness of texture information, which is then used to segment the image into uneven grids. Feature point matching pairs are then used to calculate the grid vertex homography matrix and assign different weights. Finally, a moving homography transform is used to generate the deformed target image, and a feathering fusion method is used to process the reference image and the transformed target image to generate a fused image. This method fully considers the distribution of image texture information and uses texture information to guide grid division, resulting in better alignment of overlapping image regions while also improving robustness.

[0113] Example 3

[0114] Figure 5a and Figure 5b is a set of images to be stitched, where Figure 5a is the reference image, Figure 5b is the target image; Figure 5c This is a schematic diagram of the image feature point matching pairs obtained in step 1; Figure 5dis the coarse registration image I obtained in step 2 g ; Figure 5e is the adaptive grid division result obtained by steps 3 and 4, where the contrast threshold is set to the global contrast of the coarse registration image, and the upper limit of the grid width W is max The lower limit of the grid width is 0.1 of the coarse registration image width. min is 0.01 of the width of the coarse registration image; Figure 5f is the deformed target image obtained through steps 5 and 6; Figure 5g This is the fused stitching result obtained in step 7. From the stitched image, it can be seen that compared with the traditional image stitching method with uniform grid division, the method of the present invention has certain advantages in eliminating image artifacts and improving the alignment accuracy of complex texture areas.

[0115] Example 4

[0116] Figure 6a and Figure 6b is a set of images to be stitched, where Figure 6a is the reference image, Figure 6b is the target image. Figure 6c This is a schematic diagram of the image feature point matching pairs obtained in step 1; Figure 6d is the coarse registration image I obtained in step 2 g ; Figure 6e is the adaptive grid division result obtained by steps 3 and 4, where the contrast threshold is set to the global contrast of the coarse registration image, and the upper limit of the grid width W max The lower limit of the grid width is 0.1 of the coarse registration image width. min is 0.01 of the width of the coarse registration image; Figure 6f is the deformed target image obtained through steps 5 and 6; Figure 6g This is the stitching result after fusion obtained in step 7. From the stitched image, it can be seen that the image stitched by the method of the present invention can significantly reduce the artifacts of large parallax scenes and improve the visual effect of the stitched image.

[0117] Example 5

[0118] Figure 7a-7c The proposed method is compared with the classic SIFT+RANSAC method, APAP method, AANAP method, SPHP method, and LPC method, using three metrics: root mean square error, peak signal-to-noise ratio, and structural similarity index. As can be seen from the figure, the proposed method achieves the best performance among all three image quality evaluation methods.

[0119] Example 6

[0120] The method of the present invention effectively addresses the problems of traditional grid-deformation-based image stitching methods, which ignore image texture information and produce significant variations in stitching results across different scenes. Because this method fully considers the distribution of image texture information, it is highly robust. This method effectively achieves image alignment while reducing computational complexity.

Claims

1. An adaptive grid division and image stitching method based on image texture information distribution, characterized in that: Please follow the steps below to implement it: Step 1: Extract and match the feature points of the reference image and the target image; Step 2: Calculate the global homography matrix by matching feature points, pre-register the two images, and generate a coarse registration image; Step 3: Calculate the gray level co-occurrence matrix of the coarse registration image, use contrast to represent the image texture information, and obtain the texture information of different areas of the image; Step 4: After calculating the contrast in different grid areas, compare it with the threshold, and recursively divide it to obtain the grid vertex coordinates, grid width, and grid height; Step 5: Calculate the transformation matrix of different mesh vertices based on the feature point matching pairs and the feature point weights, and deform the mesh vertices according to the transformation matrix; Step 6: Use the transformation matrix of the mesh vertices to guide the transformation of the pixels within the mesh to generate the deformed target image; Step 7: Fuse the deformed target image with the reference image and use the feathering fusion method to eliminate the image boundaries.

2. The adaptive grid division and image splicing method based on image texture information distribution according to claim 1, characterized in that: In the step 1, specifically: For the target image I r Deform it to the reference image I l The SIFT algorithm is used to extract the feature points of the two images and generate feature descriptors. The reference image I is l and target image I r The feature points in the image are matched, and then the RANSAC method is used to eliminate the mismatched pairs.

3. The adaptive grid division and image splicing method based on image texture information distribution according to claim 1, characterized in that: In the step 2, specifically: Using at least four sets of feature point matching pairs and any three sets of feature points are not collinear, the global homography transformation matrix H is calculated. The specific formula is shown in formula (1): Where: x i ′ represents the reference image I l Feature point horizontal coordinate, y i ′ represents the reference image I l Feature point vertical coordinate, x i Represents the target image I r Feature point horizontal coordinate, y i Represents the target image I r Feature point vertical coordinate, h 00 ,h 01 ,…h 22 are the 9 parameters of the homography transformation matrix H; Expanding formula (1) yields formula (2): The global homography transformation matrix is ​​calculated by solving equation (2) using the least squares method. The target image is deformed using the global homography transformation matrix and fused with the reference image to obtain the coarse registration image I g .

4. The adaptive grid division and image splicing method based on image texture information distribution according to claim 1, characterized in that: In step 3, the details are as follows: Step 3.1, calculate the coarse registration image I g Gray level co-occurrence matrix, as shown in formula (3); set the contrast threshold Th, grid width upper limit W max and the lower limit of the grid width W min ; Where: P(i,j) is the element in the gray-level co-occurrence matrix, i and j refer to the pixels with gray-level values ​​i and j respectively; C(i,j|d,q) is the number of occurrences of pixel pairs with gray-level values ​​i and j, d represents the spatial distance, θ represents the direction; N represents the number of gray-level levels; Step 3.2: The coarse registration image is set according to the upper limit of the grid width W. max Perform pre-division, and then calculate the contrast of the coarse registration image in different grid areas through the gray level co-occurrence matrix, as shown in formula (4); Among them: Contrast represents the contrast.

5. The adaptive grid division and image splicing method based on image texture information distribution according to claim 4, characterized in that: In the step 4, specifically: After calculating the contrast in different grid areas, compare it with the contrast threshold Th. If it is greater than the threshold, the current grid is divided into 4 sub-areas until the contrast in the grid is less than the threshold or the grid width is less than W. min ;Store the current grid point information, including grid vertex coordinates, grid width and grid height.

6. The adaptive grid division and image splicing method based on image texture information distribution according to claim 5, characterized in that: In the step 5, specifically: Step 5.1, assign different weights to each feature point according to the distance from the grid center point to the feature point; The weight calculation formula is shown in formula (5); Where: x * represents the center point of the grid, x i represents the characteristic point, σ is the bandwidth parameter; Weight Matrix As shown in formula (6); Step 5.2, use the feature matching point pairs to calculate the homography transformation matrix of each grid center point. The homography matrix calculation is shown in Equations (7) and (8); Where: is the coordinate of the center point of the grid after transformation, H * is the local homography transformation matrix, x * ′ is the grid center coordinate of the coarse registration image, h * Denotes the local homography matrix H * The parameterized form of Represents the Gaussian weight of the i-th feature point to the current local area, a i It is a constraint matrix constructed based on the i-th feature point matching pair, converting the homography matrix into a linear constraint.

7. The adaptive grid division and image splicing method based on image texture information distribution according to claim 6, characterized in that: In step 6, the pixel position of each pixel on the canvas after homography transformation is calculated using the homography matrices of different grids obtained in step 5; If it is inside the target image, copy the pixel on the target image to the canvas; If the mapped coordinates are not integers, they are rounded before processing.

8. The adaptive grid division and image splicing method based on image texture information distribution according to claim 1, characterized in that: In step 7, a feathering fusion algorithm is used to eliminate the image boundary; weights are assigned to pixels according to the distance from the pixel to the boundary, and the feathering fusion formula is shown in formula (9); I(x,y)=ω1(x,y)×I1(x,y)+ω2(x,y)×I2(x,y) (9) Where I(x,y) represents the grayscale value of the pixel in the boundary area of ​​the image, I1(x,y) and I2(x,y) are the grayscale values ​​of the edge pixels of the reference image and the target image on both sides of the boundary, ω1(x,y) and ω2(x,y) are weight functions and satisfy ω1(x,y)+ω2(x,y)=1; The weight is calculated based on the distance from the pixel to the boundary, as shown in formula (10); For the reference image I1, d1(x,y) is the closest distance from the pixel point (x,y) to the boundary of I1, and d2(x,y) is the closest distance from the pixel point (x,y) to the boundary of I2.

Images