A Repeated Texture Image Matching Method Based on Delaunay Triangulation

Abstract

The present invention discloses a repeated texture image matching method based on Delaunay triangulation, comprising: extracting feature points of two images by using SIFT algorithm to obtain feature vectors of each feature point; calculating the Euclidean distance between feature vectors of feature points in two images, and forming initial matching between feature points of two images according to nearest neighbors of feature vectors; selecting initial matching with good performance as seed point pairs by using local non-maximum suppression; filtering seed point pairs that do not conform to the global topological structure based on Delaunay triangulation; dividing two images into multiple circular areas with the filtered seed point pairs as the center, fitting affine transformation of local areas by PROSAC algorithm in the circular areas, and removing wrong matching. When matching images with more similar textures by using the method of the present invention, the accuracy of matching point pairs is improved. The method has a significant effect on removing wrong matching caused by repeated textures.

Description

Technical Field

[0001] The invention belongs to the technical field of image matching and relates to a repeated texture image matching method based on Delaunay triangulation. Background Art

[0002] Image matching is to find corresponding matching points in two pictures with common areas. It is the basis of many visual tasks, such as 3D reconstruction. By finding corresponding points in two images, the basic matrix between them can be obtained, so as to reconstruct the coordinates of these points in 3D space. Image matching can also be applied to image stitching. After finding the corresponding points in two pictures, they can be stitched into one picture. Image matching also plays a very important role in tasks such as image registration, remote sensing and SLAM, so it has attracted extensive research.

[0003] Existing image matching methods include simple filter-based methods, global-based methods, and local-based methods. Among the existing image matching methods, Lowe's ratio test is a simple and effective method for removing false matches. It filters out outliers by setting a threshold for the ratio of the nearest neighbor to the next nearest neighbor. Simple filters also have bidirectional matching detection. When searching for corresponding points of feature points in the first image in the second image, one matching method will be obtained. Conversely, when searching for corresponding points of feature points in the second image in the first image, another matching method may be obtained. Bidirectional matching detection retains point pairs that can be matched in both directions, and removes point pairs that cannot be matched in both directions. Simple filters can effectively remove false matches, but the results obtained by this method of removing false matches only through the feature vectors of local feature descriptors are still not very accurate.

[0004] The most widely used global-based method is the RANSAC-type robust matcher. Chinese patent CN103400388A uses the RANSAC method to eliminate false matches. The RANSAC method randomly samples from the hypothesis set, extracts the minimum sample used to generate the model, and then returns the support of the model. After reaching the maximum number of iterations, the model with the largest support is selected. The method is simple and robust, and is widely used. However, the matching point pairs finally obtained by the global RANSAC method will appear to be concentrated in texture-rich areas, and the number of matches obtained will also be relatively small. The global RANSAC algorithm also takes a long time to run when there are many false matches.

[0005] There are also some methods that eliminate false matches through local neighborhood constraints. The GMS algorithm proposed in doi:10.1109 / CVPR.2017.302. assumes that adjacent pixels in the image will move together, implements motion statistics based on the grid, improves the performance of outlier filtering, and greatly reduces the running time of the algorithm, which can be applied in real time. AdaLAM proposed in ECCV 2020, pages 770–787 integrates the best practices in a large number of mature manual method literatures in this field into a coherent framework, and achieves fast and effective outlier filtering through traditional methods. This local area-based method can achieve parallelism in each area, greatly reducing the running time, and the matching point pairs obtained are more evenly distributed in the image. Local area-based methods such as GMS and AdaLAM can make the distribution of matching point pairs more even and denser, and this method of dividing the image into several small areas that do not affect each other can also speed up the algorithm. However, in images with a lot of repeated textures (such as the appearance of a building), algorithms based on local areas will cause incorrect matches in the entire area, such as Figure 1 As shown, this is also a difficulty in image matching. Summary of the invention

[0006] To solve the above technical problems, the purpose of the present invention is to provide a repeated texture image matching method based on Delaunay triangulation, which adds global constraints to the local area based method, combines the advantages of the two, and can obtain dense and uniform matching point pairs, which has a significant effect on removing erroneous matches caused by repeated textures.

[0007] The present invention provides a repeated texture image matching method based on Delaunay triangulation, comprising the following steps:

[0008] Step 1: Use SIFT algorithm to extract feature points of two images and obtain feature vectors of each feature point;

[0009] Step 2: Calculate the Euclidean distance between the feature vectors of the feature points in the two images, and form an initial match between the feature points of the two images based on the nearest neighbors of the feature vectors;

[0010] Step 3: Use local non-maximum suppression to select initial matches with good performance as seed point pairs;

[0011] Step 4: Filter the seed point pairs that do not conform to the global topological structure based on Delaunay triangulation;

[0012] Step 5: Divide the two images into circular areas with the filtered seed point pairs as the center, and use the PROSAC algorithm to fit the affine transformation of the local area in the circular area to remove false matches.

[0013] In the repeated texture image matching method based on Delaunay triangulation of the present invention, the step 1 comprises:

[0014] Step 1.1: Perform Gaussian blur of different scales on the two images to obtain a series of images of different sizes;

[0015] Step 1.2: Find extreme points as feature points in images of different scales;

[0016] Step 1.3: Extract no more than 8000 feature points from each image, and obtain a 128-dimensional feature vector describing the feature point based on the gradient histogram in the 4×4 region around the feature point.

[0017] In the repeated texture image matching method based on Delaunay triangulation of the present invention, the step 2 is specifically: for each feature point in the first image, a feature point in the second image is selected to form an initial match, and the Euclidean distance between the two feature vectors of the two feature points of the initial match is the smallest, including the following steps:

[0018] Step 2.1: Use the distance matrix D to record the Euclidean distance between the feature vector of the feature point in the first image G1 and the feature vector of the feature point in the second image G2. The element d in the distance matrix D is ij Represents the Euclidean distance between the feature vectors of the i-th feature point in G1 and the j-th feature point in G2;

[0019] Step 2.2: For the i-th feature point in G1, find the minimum value d in the i-th row of the distance matrix D ij , then the i-th feature point in G1 and the j-th feature point in G2 are a pair of matching points;

[0020] Step 2.3: Repeat step 2.2 to find the matching points of all feature points in G1 to form several initial matching point pairs.

[0021] In the repeated texture image matching method based on Delaunay triangulation of the present invention, the step 3 is specifically as follows:

[0022] Step 3.1: Find the next nearest neighbor of each feature point in the first image, that is, the second smallest value d in each row of the distance matrix D ik ;

[0023] Step 3.2: Assign a ratio test score rt to each initially matched point pair, and generate the nearest neighbor d of the feature vector of each feature point in the first image found during the initial matching ij and the next nearest neighbor d ik The ratio of is used as the ratio test score rt of the initial matching point pair:

[0024]

[0025] Step 3.3: Use local non-maximum suppression to select the initial matching point pair with the smallest ratio test score rt in the circular local area with a radius of R1 as the seed point pair.

[0026] In the repeated texture image matching method based on Delaunay triangulation of the present invention, the step 3.3 is specifically as follows:

[0027] Step 3.3.1: Let the ratio test score of a pair of initial matching points X1 and X2 be rt1. In the circular local area with X1 as the center and R1 as the radius in the first image, if the ratio test scores of the remaining initial matching points are greater than rt1, then the initial matching point pair X1 and X2 are selected as the seed point pair;

[0028] Step 3.3.2: Calculate the radii R1 and R2 in the two images based on the image area being 120 times the area of ​​the circular local region.

[0029] In the repeated texture image matching method based on Delaunay triangulation of the present invention, the step 4 is specifically as follows:

[0030] Step 4.1: Perform Delaunay triangulation on the seed points of the first image;

[0031] Step 4.2: Connect the corresponding seed points in the second image according to the connection method of the seed points obtained by triangulation in the first image. If the second image is a triangular mesh with no intersection after connection, then all the seed point pairs are matched correctly; otherwise, proceed to step 4.3;

[0032] Step 4.3: For the seed point in the second image, count the average number of intersections between the lines connected to it and other lines, that is, the average number of intersections Where N is the number of lines connected to the seed point, and n is the number of intersections between the line connected to the seed point and other lines;

[0033] Step 4.3: Count the average number of intersections of each seed point in the second image in turn, and set a threshold. If the average number of intersections of a seed point is greater than the threshold, remove this seed point and the corresponding seed point in the first image; the threshold is the average of the average number of intersections of all seed points.

[0034] In the repeated texture image matching method based on Delaunay triangulation of the present invention, the step 5 is specifically as follows:

[0035] Step 5.1: Divide the two images into circular areas with the filtered seed point pairs as the center;

[0036] Step 5.2: Sort the ratio test scores of the initial matching point pairs in each circular area divided in step 5.1 from small to large;

[0037] Step 5.3: Each initial matching point pair within the circular area conforms to the central affine transformation, the formula of which is:

[0038]

[0039] Among them, x, y and x', y' represent the coordinate values ​​of the initial matching point pairs in the first image and the second image. It is an affine matrix. Two equality constraints are obtained through a set of initial matching point pairs. An affine matrix is ​​calculated from two sets of initial matching point pairs.

[0040] Step 5.4: Use the PROSAC algorithm to select the top two initial matching point pairs in each circular area in the first iteration. In the second iteration, randomly select two point pairs from the top three initial matching point pairs in each circular area. In the third iteration, randomly select two point pairs from the top four initial matching point pairs. Select points in the increasing sampling space in turn. For each circular area, use the two point pairs selected in each iteration to calculate the affine matrix A corresponding to the current number of iterations. i , i represents the number of iterations;

[0041] Step 5.5: According to r i =||A i [x,y] T -[x',y'] T ||Calculate the error r of all initial matching point pairs except two sampling points in each circular area at each iteration i , the initial matching point pairs with errors less than the error threshold are inliers, and those with errors greater than the error threshold are outliers. The number of inliers in each circular area of ​​each iteration is recorded. After 128 iterations, the affine matrix calculated from the iteration with the largest number of inliers is selected as the affine matrix of the circular area;

[0042] Step 5.6: For each circular region, remove the outer points of the affine matrix that do not conform to the circular region.

[0043] In the repeated texture image matching method based on Delaunay triangulation of the present invention, the step 5.1 is specifically as follows:

[0044] Step 5.1.1: Let R'1 be 3 times R1, and R'2 be 3 times R2;

[0045] Step 5.1.2: Select a series of seed point pairs through step 3.3, and divide the first image and the second image into circular areas with the seed points in these seed point pairs as the center and R'1 and R'2 as the radius respectively.

[0046] The present invention discloses a repeated texture image matching method based on Delaunay triangulation. The Delaunay triangulation algorithm is used to filter seed points that do not conform to the global topological structure, and global constraints are added to the method based on the local area, thereby combining the advantages of the two. The method has a significant improvement effect on the situation where the entire area is mismatched when there are many similar textures in the image, and the accuracy of the matching point pairs finally obtained is improved. Experiments show that the method of the present invention can finally obtain dense and uniform matching point pairs, and has a significant effect on removing mismatches caused by repeated textures. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a flow chart of a repeated texture image matching method based on Delaunay triangulation of the present invention;

[0048] Figure 2a It is the triangular network diagram of the first picture;

[0049] Figure 2b It is the triangular network diagram without intersection points in the second picture;

[0050] Figure 2c It is the triangulated network diagram with intersection points of the second picture;

[0051] Figure 3 This is a schematic diagram of the local neighborhood division centered on the seed point. DETAILED DESCRIPTION

[0052] The global image matching method takes a long time to calculate, and the matching point pairs obtained are unevenly distributed in the image. Image matching based on local areas can widely cover various areas of the image and obtain abundant matching point pairs. However, when the image contains a large number of similar textures, it is easy to produce regional mismatches only through local relationships. In order to obtain a large number of uniform matching point pairs while ensuring the accuracy of matching, the present invention combines global and local constraints, and optimizes the seed points through Delaunay triangulation based on the invariance of the scene structure. First, the seed points in the first image are Delaunay triangulated to form a triangular mesh, and then the corresponding points in the second image are also connected according to the connection method of the seed points in the first image. If all the matching point pairs are correct, a triangular mesh without intersections can still be formed after the connection, otherwise there are errors. The present invention filters the seed points by setting a threshold for the average number of intersections, which effectively reduces the situation of regional mismatches caused by a large number of similar textures, making the final matching more accurate.

[0053] A repeated texture image matching method based on Delaunay triangulation of the present invention is described in detail below with reference to the accompanying drawings.

[0054] like Figure 1 As shown, a repeated texture image matching method based on Delaunay triangulation of the present invention comprises the following steps:

[0055] Step 1: Using SIFT algorithm to extract feature points of two images and obtain feature vectors of each feature point, the step 1 includes:

[0056] Step 1.1: Perform Gaussian blur of different scales on the two images to obtain a series of images of different sizes;

[0057] Step 1.2: Find extreme points as feature points in images of different scales;

[0058] Step 1.3: Extract no more than 8,000 feature points from each image, and obtain a 128-dimensional feature vector describing the feature point based on the gradient histogram in the 4×4 region around the feature point (each histogram has 8 directions).

[0059] Step 2: Calculate the Euclidean distance between the feature vectors of the feature points in the two images, and form an initial match between the feature points of the two images based on the nearest neighbors of the feature vectors;

[0060] The step 2 is specifically: for each feature point in the first image, a feature point in the second image is selected to form an initial match, and the Euclidean distance between the two feature vectors of the two feature points of the initial match is the smallest, including the following steps:

[0061] Step 2.1: Use the distance matrix D to record the Euclidean distance between the feature vector of the feature point in the first image G1 and the feature vector of the feature point in the second image G2. The element d in the distance matrix D is ij Represents the Euclidean distance between the feature vectors of the i-th feature point in G1 and the j-th feature point in G2;

[0062] Step 2.2: For the i-th feature point in G1, find the minimum value d in the i-th row of the distance matrix D ij , then the i-th feature point in G1 and the j-th feature point in G2 are a pair of matching points;

[0063] Step 2.3: Repeat step 2.2 to find the matching points of all feature points in G1 to form several initial matching point pairs.

[0064] The initial match obtained in step 2 is found based on the nearest neighbor of the feature vector, and the assumed match formed only by the local descriptor is very inaccurate, so the assumed match needs to be further filtered.

[0065] Step 3: Use local non-maximum suppression to select initial matches with good performance as seed point pairs;

[0066] In order to retain the advantages of fast calculation speed and uniform distribution of seed points based on local methods, a method based on local areas is first used. When dividing the neighborhood, local non-maximum suppression is first used to select seed points, and then the neighborhood is divided near the seed points. To perform local non-maximum suppression, each initial matching point pair must first be assigned a score. Here, the value of the ratio test is used as the score of the matching point pair. The step 3 is specifically as follows:

[0067] Step 3.1: Find the next nearest neighbor of each feature point in the first image, that is, the second smallest value d in each row of the distance matrix D ik ;

[0068] Step 3.2: Assign a ratio test score rt to each initially matched point pair, and generate the nearest neighbor d of the feature vector of each feature point in the first image found during the initial matching ij and the next nearest neighbor d ik The ratio of is used as the ratio test score rt of the initial matching point pair:

[0069]

[0070] Step 3.3: Use local non-maximum suppression to select the initial matching point pair with the smallest ratio test score rt in the circular local area with a radius of R1 as the seed point pair, specifically:

[0071] Step 3.3.1: Let the ratio test score of a pair of initial matching points X1 and X2 be rt1. In the circular local area with X1 as the center and R1 as the radius in the first image, if the ratio test scores of the remaining initial matching points are greater than rt1, then the initial matching point pair X1 and X2 are selected as the seed point pair;

[0072] Step 3.3.2: Calculate the radii R1 and R2 in the two images based on the image area being 120 times the area of ​​the circular local region.

[0073] Step 4: Filter the seed point pairs that do not conform to the global topological structure based on Delaunay triangulation;

[0074] In step 3, a series of seed point pairs are first selected. The next step is to eliminate false matches in the local neighborhood. This method can parallelize the operation of each neighborhood, greatly shorten the running time, and the final matches are more evenly distributed in the image. However, only the local neighborhood-based method will produce false matches in the entire area in images with many repeated textures. In order to improve this situation, constraints based on the global topology structure are added to filter the seed points.

[0075] When the same scene is photographed from two different angles, the topological structure of the scene is stable and will not change with the change of the shooting angle. For example, the relative position of each point of the face structure remains unchanged no matter from which direction it is photographed. The ears are always on both sides of the eyes and will not move to the middle of the eyes. Therefore, we can further constrain the seed points through the global topological distribution. The present invention uses Delaunay triangulation to determine the topological distribution of the seed points in the two images.

[0076] Triangulation is to connect the point set on the plane with closed line segments to form a triangulated network, and the connected figures have no intersecting edges, and the union of triangles is the convex hull of the scattered point set. There are many results of triangulating the point set on the plane. The more symmetrical the triangular shape is, the better the subsequent image processing effect will be. In order to make the result unique and make the triangular shape more symmetrical, this paper uses Delaunay triangulation, which is a special triangulation. Delaunay triangulation satisfies the empty circle property. Any four points cannot be in the same circle, and the triangular mesh formed is unique. The step 4 is specifically as follows:

[0077] Step 4.1: Perform Delaunay triangulation on the seed points of the first image to obtain Figure 2a The triangulated network shown;

[0078] Step 4.2: Connect the corresponding seed points in the second image according to the connection method of the seed points obtained by triangulation in the first image. If the second image is a triangular mesh with no intersection after connection, such as Figure 2b As shown in , all seed points are matched correctly; if there are line segments crossing after connection, it means that there are mismatched points, such as Figure 2c As shown, otherwise, execute step 4.3;

[0079] Step 4.3: For the seed point in the second image, count the average number of intersections between the lines connected to it and other lines, that is, the average number of intersections Where N is the number of lines connected to the seed point, and n is the number of intersections between the line connected to the seed point and other lines;

[0080] Step 4.3: Count the average number of intersections of each seed point in the second image in turn, and set a threshold. If the average number of intersections of a seed point is greater than the threshold, remove this seed point and the corresponding seed point in the first image; the threshold is the average of the average number of intersections of all seed points.

[0081] By combining the global constraint with the local area-based mismatch filtering through step 4, the erroneous seed points can be initially filtered, thereby reducing the subsequent mismatching of the entire neighborhood.

[0082] Step 5: Divide the two images into circular areas with the filtered seed point pairs as the center, and use the PROSAC algorithm to fit the affine transformation of the local area in the circular area to remove the wrong matching. The specific step 5 is as follows:

[0083] Step 5.1: Divide the two images into circular areas with the filtered seed point pairs as the center, specifically:

[0084] Step 5.1.1: Let R'1 be 3 times R1, and R'2 be 3 times R2;

[0085] Step 5.1.2: Select a series of seed point pairs through step 3.3, and divide the first image and the second image into circular areas with the seed points in these seed point pairs as the center and R'1 and R'2 as the radius, respectively, as follows: Figure 3 shown.

[0086] Step 5.2: Sort the ratio test scores of the initial matching point pairs in each circular area divided in step 5.1 from small to large;

[0087] Step 5.3: Each initial matching point pair within the circular area conforms to the central affine transformation, the formula of which is:

[0088]

[0089] Among them, x, y and x', y' represent the coordinate values ​​of the initial matching point pairs in the first image and the second image. It is an affine matrix. Two equality constraints are obtained through a set of initial matching point pairs. An affine matrix is ​​calculated from two sets of initial matching point pairs.

[0090] Step 5.4: Use the PROSAC algorithm to select the top two initial matching point pairs in each circular area in the first iteration. In the second iteration, randomly select two point pairs from the top three initial matching point pairs in each circular area. In the third iteration, randomly select two point pairs from the top four initial matching point pairs. Select points in the increasing sampling space in turn. For each circular area, use the two point pairs selected in each iteration to calculate the affine matrix A corresponding to the current number of iterations. i , i represents the number of iterations;

[0091] Step 5.5: According to r i =||A i [x,y] T -[x',y'] T ||Calculate the error r of all initial matching point pairs except two sampling points in each circular area at each iteration i , the initial matching point pairs with errors less than the error threshold are inliers, and those with errors greater than the error threshold are outliers. The number of inliers in each circular area of ​​each iteration is recorded. After 128 iterations, the affine matrix calculated from the iteration with the largest number of inliers is selected as the affine matrix of the circular area;

[0092] Step 5.6: For each circular region, remove the outer points of the affine matrix that do not conform to the circular region.

[0093] Similar textures are a difficult point in image matching. The present invention adds global constraints to PROSAC based on local regions, and applies Delaunay triangulation to filter seed points that do not conform to the global topological structure, thereby avoiding the formation of regional mismatches. Through these improvements, the situation of mismatching of the entire region when there are many similar textures in the image is improved, and the accuracy of the matching point pairs finally obtained is improved. The method of the present invention can finally obtain dense and uniform matching point pairs, which has a significant effect on removing mismatches caused by repeated textures.

[0094] The above description is only a preferred embodiment of the present invention and is not intended to limit the concept of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A repeated texture image matching method based on Delaunay triangulation, characterized in that: The steps include: Step 1: Use SIFT algorithm to extract feature points of two images and obtain feature vectors of each feature point; Step 2: Calculate the Euclidean distance between the feature vectors of the feature points in the two images, and form an initial match between the feature points of the two images based on the nearest neighbors of the feature vectors; Step 3: Use local non-maximum suppression to select initial matches with good performance as seed point pairs; Step 4: Filter the seed point pairs that do not conform to the global topological structure based on Delaunay triangulation, specifically: Step 4.1: Perform Delaunay triangulation on the seed points of the first image; Step 4.2: Connect the corresponding seed points in the second image according to the connection method of the seed points obtained by triangulation in the first image. If the second image is a triangular mesh with no intersections after connection, then all the seed point pairs are matched correctly; Otherwise, proceed to step 4.3; Step 4.3: For the seed point in the second image, count the average number of intersections between the lines connected to it and other lines, that is, the average number of intersections Where N is the number of lines connected to the seed point, and n is the number of intersections between the line connected to the seed point and other lines; Step 4.4: Count the average number of intersections of each seed point in the second image in turn, and set a threshold. If the average number of intersections of a seed point is greater than the threshold, remove this seed point and the seed point corresponding to it in the first image; the threshold is the average of the average number of intersections of all seed points; Step 5: Divide the two images into circular areas with the filtered seed point pairs as the center, and use the PROSAC algorithm to fit the affine transformation of the local area in the circular area to remove false matches.

2. The repeated texture image matching method based on Delaunay triangulation as claimed in claim 1, characterized in that: The step 1 comprises: Step 1.1: Perform Gaussian blur of different scales on the two images to obtain a series of images of different sizes; Step 1.2: Find extreme points as feature points in images of different scales; Step 1.3: Extract no more than 8000 feature points from each image, and obtain a 128-dimensional feature vector describing the feature point based on the gradient histogram in the 4×4 region around the feature point.

3. The repeated texture image matching method based on Delaunay triangulation as claimed in claim 1, characterized in that: The step 2 is specifically: for each feature point in the first image, a feature point in the second image is selected to form an initial match, and the Euclidean distance between the two feature vectors of the two feature points of the initial match is the smallest, including the following steps: Step 2.1: Use the distance matrix D to record the Euclidean distance between the feature vector of the feature point in the first image G1 and the feature vector of the feature point in the second image G2. The element d in the distance matrix D is ij Represents the Euclidean distance between the feature vectors of the i-th feature point in G1 and the j-th feature point in G2; Step 2.2: For the i-th feature point in G1, find the minimum value d in the i-th row of the distance matrix D ij , then the i-th feature point in G1 and the j-th feature point in G2 are a pair of matching points; Step 2.3: Repeat step 2.2 to find the matching points of all feature points in G1 to form several initial matching point pairs.

4. The repeated texture image matching method based on Delaunay triangulation as claimed in claim 3, characterized in that: The step 3 is specifically as follows: Step 3.1: Find the next nearest neighbor of each feature point in the first image, that is, the second smallest value d in each row of the distance matrix D ik ; Step 3.2: Assign a ratio test score rt to each initially matched point pair, and generate the nearest neighbor d of the feature vector of each feature point in the first image found during the initial matching ij and the next nearest neighbor d ik The ratio of is used as the ratio test score rt of the initial matching point pair: Step 3.3: Use local non-maximum suppression to select the initial matching point pair with the smallest ratio test score rt in the circular local area with a radius of R1 as the seed point pair.

5. The repeated texture image matching method based on Delaunay triangulation as claimed in claim 4, characterized in that: The step 3.3 is specifically as follows: Step 3.3.1: Let the ratio test score of a pair of initial matching points X1 and X2 be rt1. In the circular local area with X1 as the center and R1 as the radius in the first image, if the ratio test scores of the remaining initial matching points are greater than rt1, then the initial matching point pair X1 and X2 are selected as the seed point pair; Step 3.3.2: Calculate the radii R1 and R2 in the two images based on the image area being 120 times the area of ​​the circular local region.

6. The repeated texture image matching method based on Delaunay triangulation as claimed in claim 4, characterized in that: The step 5 is specifically as follows: Step 5.1: Divide the two images into circular areas with the filtered seed point pairs as the center; Step 5.2: Sort the ratio test scores of the initial matching point pairs in each circular area divided in step 5.1 from small to large; Step 5.3: Each initial matching point pair within the circular area conforms to the central affine transformation, the formula of which is: Among them, x, y and x', y' represent the coordinate values ​​of the initial matching point pairs in the first image and the second image. It is an affine matrix. Two equality constraints are obtained through a set of initial matching point pairs. An affine matrix is ​​calculated from two sets of initial matching point pairs. Step 5.4: Use the PROSAC algorithm to select the top two initial matching point pairs in each circular area in the first iteration. In the second iteration, randomly select two point pairs from the top three initial matching point pairs in each circular area. In the third iteration, randomly select two point pairs from the top four initial matching point pairs. Select points in the increasing sampling space in turn. For each circular area, use the two point pairs selected in each iteration to calculate the affine matrix A corresponding to the current number of iterations. i , i represents the number of iterations; Step 5.5: According to Calculate the error r of all initial matching point pairs except two sampling points in each circular area at each iteration i , the initial matching point pairs with errors less than the error threshold are inliers, and those with errors greater than the error threshold are outliers. The number of inliers in each circular area of ​​each iteration is recorded. After 128 iterations, the affine matrix calculated from the iteration with the largest number of inliers is selected as the affine matrix of the circular area; Step 5.6: For each circular region, remove the outer points of the affine matrix that do not conform to the circular region.

7. The repeated texture image matching method based on Delaunay triangulation as claimed in claim 6, characterized in that: The step 5.1 is specifically as follows: Step 5.1.1: Let R1' be 3 times R1, and R'2 be 3 times R2; Step 5.1.2: Select a series of seed point pairs through step 3.3, and divide the first image and the second image into circular areas with the seed points in these seed point pairs as the center and R1' and R'2 as the radius respectively.

Images