Field image mosaic method based on Delauny algorithm and Kruskal maximum spanning tree algorithm

Abstract

The present invention relates to a field image stitching method based on the Delauny algorithm and the Kruskal maximum spanning tree algorithm, comprising the following steps: obtaining unmanned aerial vehicle (UAV) field images and position information corresponding to the UAV field images; reprojecting each UAV field image to obtain an initial projection matrix; constructing a triangulation network using the Delauny algorithm; matching the features of images corresponding to every two nodes using the SIFT operator; generating a stitching sequence; obtaining a panoramic image; calculating the global registration error of the panoramic image, and optimizing the global registration error of the panoramic image. In view of the fact that feature point matching of UAV field images is difficult and prone to mismatching, the present invention uses a mutual correlation coefficient to detect feature point mismatching; the present invention uses the Delauny algorithm to construct a triangulation network, which can greatly reduce the number of local registrations; and the present invention can be used with a large number of UAV images.

Description

Technical Field

[0001] The invention relates to the technical fields of photogrammetry and image processing, in particular to a field image splicing method based on a Delauny algorithm and a Kruskal maximum spanning tree algorithm. Background Art

[0002] In agricultural production, it is crucial to accurately obtain phenotypic information of field crops, such as chlorophyll content, biomass, and crop yield. In the field, manually operating instruments to obtain some phenotypic information is time-consuming and labor-intensive, extremely inefficient, and it is impossible to obtain crop phenotypic information in a short period of time. Timely and accurate estimation of crop phenotypic information can reflect the growth status of crops and enable timely implementation of corresponding measures, which is of positive significance for improving crop yield, drought resistance, and lodging resistance. With the rapid development of remote sensing satellites, UAV platforms, and professional crop phenotyping platforms, it has become an important method to acquire field data through phenotyping platforms carrying sensors such as visible light cameras, multispectral cameras, and lidar. Among them, stitching images obtained by visible light or multispectral cameras is essential.

[0003] Remote sensing satellites are increasingly becoming a platform for acquiring phenotypic parameters. They offer advantages such as high temporal resolution and efficiency, making them particularly effective for crops covering large areas. However, remote sensing images have low spatial resolution and are prone to image aliasing. Unmanned aerial vehicle (UAV) platforms, which capture field images, offer agile takeoff and landing, ease of operation, and higher spatial resolution than remote sensing platforms. However, they face challenges such as limited image coverage and a high number of images, necessitating the stitching of multiple images to visualize the entire captured area. Therefore, image stitching algorithms are a key technology for acquiring phenotypic information about field crops.

[0004] Field imagery is characterized by monotonous features and the tendency for feature points to be mismatched. Current research uses relatively rich drone imagery, making feature point matching easy. However, the number of images involved in stitching is small, and no analysis of the characteristics of field imagery is conducted. Alternatively, only two drone-acquired field images are stitched together, without considering the possibility of stitching multiple field images together in a panoramic view. Summary of the Invention

[0005] In order to solve the problem that field images are easily mismatched and too few participating images, the purpose of the present invention is to provide a field image stitching method based on the Delauny algorithm and the Kruskal maximum spanning tree algorithm, which can detect feature point mismatches, greatly reduce the number of local alignments, and can participate in a large number of drone images.

[0006] To achieve the above object, the present invention adopts the following technical solution: a field image stitching method based on Delauny algorithm and Kruskal maximum spanning tree algorithm, the method comprising the following steps in order:

[0007] (1) Obtaining UAV field images and the pose information corresponding to the UAV field images, the pose information of each UAV field image includes internal orientation elements and external orientation elements; the internal orientation elements include the internal parameter matrix, focal length and distortion parameters, and the external orientation elements include X s 、Y s 、Z s 、ω、 and γ, where X s 、Y s 、Z s is the coordinate of the photography center, ω, γ is the azimuth angle;

[0008] (2) Reproject each UAV field image according to its interior and exterior orientation elements to obtain the initial projection matrix;

[0009] (3) The photographic center coordinates X of the drone field image s 、Y s Construct nodes and form a triangulation network using the Delauny algorithm;

[0010] (4) Through the nodes in the triangulated network, the images corresponding to each two nodes are matched with feature points using the SIFT operator, and then the homography matrix is ​​obtained through the initial projection matrix and projected. If the cross-correlation coefficient after projection is greater than 0.5 or less than -0.5, the cross-correlation coefficient and the corresponding feature point matching results are retained; otherwise, the cross-correlation coefficient and the corresponding feature point matching results are deleted;

[0011] (5) The mutual correlation coefficients retained in step (4) are used as weights, and the image numbers corresponding to the retained mutual correlation coefficients are used as nodes to form an undirected graph, and the splicing order is generated by the Kruskal maximum spanning tree algorithm and the DFS traversal algorithm;

[0012] (6) Taking the first node in the splicing order as the reference of the panorama, projecting according to the splicing order and the matching results of the retained feature points to obtain the panorama;

[0013] (7) Calculate the global registration error of the panorama and optimize the global registration error of the panorama.

[0014] The step (2) specifically includes the following steps:

[0015] (2a) First, the UAV field image is dedistorted according to the intrinsic parameter matrix and distortion parameters, as shown in formula (1), and the dedistorted image is saved;

[0016]

[0017] Among them, (x, y) is the pixel coordinate, r 2 =x 2 +y 2 , k1, k2, k3 are radial distortion parameters, p1, p2 are tangential distortion parameters, x distortion and y distortion is the distortion amount;

[0018] (2b) According to the exterior orientation elements of the drone image and the focal length of the drone camera, the WGS84 coordinates of the four corners of the image are obtained, namely the coordinates of the four corner points (x1, y1), (x2, y2), (x3, y3) and (x4, y4). The collinearity principle is used to obtain the collinear coordinates (X1, Y1, Z4) of the four corner points at the average ground elevation. avr )、(X2,Y2,Z avr )、(X3,Y3,Z avr ) and (X4, Y4, Z avr ):

[0019]

[0020] Among them, Z avr is the average elevation, f is the focal length, λ is the scaling factor,

[0021] (2c) Based on the coordinates of the four corner points of the image and the collinear coordinates obtained in step (2b), calculate the corresponding homography matrix and save it as the initial projection matrix:

[0022]

[0023] Among them, the initial projection matrix h0 to h7 are elements of the initial projection matrix H.

[0024] The step (3) specifically refers to: the coordinates of each image [X s ,Y s ] as nodes, and generate a triangulation network through the Delauny algorithm. If there are edges between the nodes in the network, it means that the images corresponding to the two nodes may have overlapping parts, and finally save it in the format [node, node].

[0025] The step (4) specifically includes the following steps:

[0026] (4a) Using the triangulated network obtained in step (3), the images corresponding to each two nodes i and j are used to extract their respective feature points and feature vectors using the SIFT operator. It is assumed that the projection image i needs to be projected into the coordinate system of the reference image j using the homography matrix;

[0027] (4b) Perform ratio coarse matching and RANSAC algorithm fine matching on the extracted eigenvectors, and then calculate the homography matrix H i,j ;Through the obtained homography matrix H i,j , get the homography matrix that projects the projection image into the reference image coordinate system:

[0028] H i =H i,j H j (4)

[0029] Among them, H i and H j are the initial projection matrices of images i and j respectively;

[0030] (4c) Calculate the mutual correlation coefficient of the pixels in the overlapping area, as shown in formula (5). If the mutual correlation coefficient is greater than 0, it indicates that the overlapping similarity is positively correlated, and the closer it is to 1, the higher the correlation. If the mutual correlation coefficient is less than 0, it indicates that the overlapping areas are negatively correlated, and the closer it is to -1, the higher the correlation. If the mutual correlation coefficient is equal to 0, it indicates that the overlapping areas are uncorrelated:

[0031]

[0032] Where ncc represents the cross-correlation coefficient, m represents the number of pixels in the overlapping area, and I i,p Indicates the grayscale of the pth pixel in the overlapping area of ​​image i, I j,p represents the grayscale of the p-th pixel in the overlapping area of ​​image j; represents the average grayscale value of pixels in the overlapping area of ​​image i, represents the average grayscale value of pixels in the overlapping area of ​​image j;

[0033] (4d) If the mutual correlation coefficient is less than 0.5 and greater than -0.5, it indicates that the registration effect is too poor and feature point mismatching is likely to occur. The feature point matching results corresponding to the mutual correlation coefficient less than 0.5 and greater than -0.5 are deleted, and the remaining feature point matching results are saved.

[0034] The step (5) specifically includes the following steps:

[0035] (5a) The mutual correlation coefficient retained in step (4) is magnified to 100 times, and the absolute value is taken as the weight E. The image sequence number corresponding to the retained mutual correlation coefficient is used as the node V to form an undirected graph G =<V,E> ;

[0036] (5b) Generate a maximum spanning tree using the Kruskal maximum spanning tree algorithm, taking the node of the edge with the largest weight as the starting node, and generate the image splicing order using the DFS traversal algorithm.

[0037] The step (6) specifically includes the following steps:

[0038] (6a) Taking the first image in the stitching order as the reference of the panorama, the initial projection matrix is ​​used as the homography matrix of the first image;

[0039] (6b) Match feature points one by one using the SIFT operator in the stitching order, record the homography matrix of the image, and perform image fusion on the overlapping areas. Repeat step (6b) until all images are stitched together to obtain a panoramic image.

[0040] The step (7) specifically includes the following steps:

[0041] (7a) After the splicing is completed, the error between the nodes of the undirected graph is calculated:

[0042]

[0043] Among them, error i,j represents the local registration error between image i and image j, n is the number of feature point matches between image i and image j; (x i,k ,y i,k ) is the kth feature point of image i, (x j,k ,y j,k ) is the kth feature point of image j, H i (x i,k ,y i,k ) is the kth feature point of image i through the uniqueness matrix H i Transform to the coordinates in the panorama, H j (x j,k ,y j,k ) is the uniqueness matrix H of the kth feature point of image j j Transform to coordinates in the panorama;

[0044] (7b) Calculate the global registration error:

[0045]

[0046] Among them, error represents the global registration error, where i∈[0,s], j∈[0,t];

[0047] (7c) using the Nelder-Mead algorithm to optimize the global registration error and save the homography matrix of each image transformed into the panoramic image after optimization;

[0048] (7d) Transform the image into a panorama according to the saved homography matrix.

[0049] It can be seen from the above technical solution that the beneficial effects of the present invention are: first, in view of the situation that the feature points of field drone images are difficult to match and prone to mismatching, the present invention uses the mutual correlation coefficient to check the mismatching of feature points; second, the present invention uses the Delauny algorithm to construct a triangulation network, which can greatly reduce the number of local alignments; third, the present invention can involve a large number of drone images; fourth, the present invention solves the shortcomings of field images being prone to mismatching and too few participating images by taking 132 drone field images as research objects; fifth, feature point mismatching will lead to errors in the calculation of the homography matrix, which will have a significant impact on the calculation of the homography matrix of subsequent images. The present invention uses the mutual correlation coefficient to check whether the feature points are mismatched, and eliminates feature point pairs with poor alignment effects; sixth, the present invention uses the Nelder-Mead algorithm to optimize the alignment results, which can reduce the global error to a certain extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flow chart of the method of the present invention;

[0051] Figure 2 Schematic diagram of the triangulation network constructed by the Delauny algorithm;

[0052] Figure 3 Schematic diagram of the result of local image registration between images in the present invention;

[0053] Figure 4 This is a schematic diagram of the result of removing some weights in the present invention;

[0054] Figure 5 This is a schematic diagram of the splicing sequence generated in the present invention;

[0055] Figure 6 It is a panoramic view in the present invention;

[0056] Figure 7 This is the panoramic view optimized by the present invention. DETAILED DESCRIPTION

[0057] like Figure 1 As shown, a field image stitching method based on the Delauny algorithm and the Kruskal maximum spanning tree algorithm includes the following steps in sequence:

[0058] (1) Obtaining UAV field images and the pose information corresponding to the UAV field images, the pose information of each UAV field image includes internal orientation elements and external orientation elements; the internal orientation elements include the internal parameter matrix, focal length and distortion parameters, and the external orientation elements include X s 、Y s 、Z s 、ω、 and γ, where X s 、Y s 、Z s is the coordinate of the photography center, ω, γ is the azimuth angle;

[0059] (2) Reproject each UAV field image according to its interior and exterior orientation elements to obtain the initial projection matrix;

[0060] (3) The photographic center coordinates X of the drone field image s 、Y s Construct nodes and form a triangulation network through the Delauny algorithm, namely the Delaunay triangulation algorithm;

[0061] (4) Through the nodes in the triangulated network, the images corresponding to each two nodes are matched with feature points using the SIFT operator, and then the homography matrix is ​​obtained through the initial projection matrix and projected. If the cross-correlation coefficient after projection is greater than 0.5 or less than -0.5, the cross-correlation coefficient and the corresponding feature point matching results are retained; otherwise, the cross-correlation coefficient and the corresponding feature point matching results are deleted;

[0062] (5) The mutual correlation coefficients retained in step (4) are used as weights, and the image numbers corresponding to the retained mutual correlation coefficients are used as nodes to form an undirected graph, and the splicing order is generated by the Kruskal maximum spanning tree algorithm (Kruskal algorithm) and the DFS traversal algorithm (DFS traversal algorithm);

[0063] (6) Taking the first node in the splicing order as the reference of the panorama, projecting according to the splicing order and the matching results of the retained feature points to obtain the panorama;

[0064] (7) Calculate the global registration error of the panorama and optimize the global registration error of the panorama.

[0065] The step (2) specifically includes the following steps:

[0066] (2a) First, the UAV field image is dedistorted according to the intrinsic parameter matrix and distortion parameters, as shown in formula (1), and the dedistorted image is saved;

[0067]

[0068] Among them, (x, y) is the pixel coordinate, r 2 =x 2 +y 2 , k1, k2, k3 are radial distortion parameters, p1, p2 are tangential distortion parameters, x distortion and y distortion is the distortion amount;

[0069] (2b) According to the exterior orientation elements of the drone image and the focal length of the drone camera, the WGS84 coordinates of the four corners of the image are obtained, namely the coordinates of the four corner points (x1, y1), (x2, y2), (x3, y3) and (x4, y4). The collinearity principle is used to obtain the collinear coordinates (X1, Y1, Z4) of the four corner points at the average ground elevation. avr )、(X2,Y2,Z avr )、(X3,Y3,Z avr ) and (X4, Y4, Z avr ):

[0070]

[0071] Among them, Z avr is the average elevation, f is the focal length, λ is the scaling factor,

[0072] (2c) Based on the coordinates of the four corner points of the image and the collinear coordinates obtained in step (2b), calculate the corresponding homography matrix and save it as the initial projection matrix:

[0073]

[0074] Among them, the initial projection matrix h0 to h7 are elements of the initial projection matrix H.

[0075] like Figure 2 As shown, the step (3) specifically refers to: the coordinates of each image [X s ,Y s ] as nodes, and generate a triangulation network through the Delauny algorithm. If there are edges between the nodes in the network, it means that the images corresponding to the two nodes may have overlapping parts, and finally save it in the format [node, node].

[0076] The step (4) specifically includes the following steps:

[0077] (4a) Using the triangulated network obtained in step (3), the images corresponding to each two nodes i and j are used to extract their respective feature points and feature vectors using the SIFT operator. It is assumed that the projection image i needs to be projected into the coordinate system of the reference image j using the homography matrix;

[0078] (4b) Perform ratio coarse matching and RANSAC algorithm fine matching on the extracted eigenvectors, and then calculate the homography matrix H i,j ;Through the obtained homography matrix H i,j , get the homography matrix that projects the projection image into the reference image coordinate system:

[0079] H i =H i,j H j (4)

[0080] Among them, H i and H j are the initial projection matrices of images i and j respectively;

[0081] (4c) Calculate the mutual correlation coefficient of the pixels in the overlapping area, as shown in formula (5). If the mutual correlation coefficient is greater than 0, it indicates that the overlapping similarity is positively correlated, and the closer it is to 1, the higher the correlation. If the mutual correlation coefficient is less than 0, it indicates that the overlapping areas are negatively correlated, and the closer it is to -1, the higher the correlation. If the mutual correlation coefficient is equal to 0, it indicates that the overlapping areas are uncorrelated:

[0082]

[0083] Where ncc represents the cross-correlation coefficient, m represents the number of pixels in the overlapping area, and I i,p Indicates the grayscale of the pth pixel in the overlapping area of ​​image i, I j,p represents the grayscale of the p-th pixel in the overlapping area of ​​image j; represents the average grayscale value of pixels in the overlapping area of ​​image i, represents the average grayscale value of pixels in the overlapping area of ​​image j; Figure 3 The result of local image registration between the images in step (4c) is obtained by multiplying the absolute value of the mutual correlation coefficient by 100 times.

[0084] (4d) If the mutual correlation coefficient is less than 0.5 and greater than -0.5, it indicates that the registration effect is too poor and feature point mismatching is likely to occur. The feature point matching results corresponding to the mutual correlation coefficient less than 0.5 and greater than -0.5 are deleted, and the remaining feature point matching results are saved.

[0085] The step (5) specifically includes the following steps:

[0086] (5a) The mutual correlation coefficient retained in step (4) is magnified to 100 times, and the absolute value is taken as the weight E. The image sequence number corresponding to the retained mutual correlation coefficient is used as the node V to form an undirected graph G =<V,E> ; Figure 4 is the result of removing some weights in step (5a).

[0087] (5b) Generate a maximum spanning tree using the Kruskal maximum spanning tree algorithm, taking the node of the edge with the largest weight as the starting node, and generate the image splicing order using the DFS traversal algorithm. Figure 5 The stitching order generated in step (5b) is shown in the figure. The color of the brackets indicates the registration result of the two images in step (4c). The numbers in the brackets indicate the stitching order of the images in the global registration. The arrows indicate the coordinate system of the projected image transformed into the reference image through homography.

[0088] The step (6) specifically includes the following steps:

[0089] (6a) Taking the first image in the stitching order as the reference of the panorama, the initial projection matrix is ​​used as the homography matrix of the first image;

[0090] (6b) Match the feature points one by one using the SIFT operator in the stitching order, record the homography matrix of the image, and perform image fusion on the overlapping areas. Repeat step (6b) until all images are stitched together to obtain a panoramic image, as shown in the following example. Figure 6 shown.

[0091] The step (7) specifically includes the following steps:

[0092] (7a) After the splicing is completed, the error between the nodes of the undirected graph is calculated:

[0093]

[0094] Among them, error i,j represents the local registration error between image i and image j, n is the number of feature point matches between image i and image j; (x i,k ,y i,k ) is the kth feature point of image i, (x j,k ,y j,k ) is the kth feature point of image j, H i (x i,k ,y i,k ) is the kth feature point of image i through the uniqueness matrix H i Transform to the coordinates in the panorama, H j (x j,k ,y j,k ) is the uniqueness matrix H of the kth feature point of image j j Transform to coordinates in the panorama;

[0095] (7b) Calculate the global registration error:

[0096]

[0097] Among them, error represents the global registration error, where i∈[0,s], j∈[0,t];

[0098] (7c) using the Nelder-Mead algorithm to optimize the global registration error and save the homography matrix of each image transformed into the panoramic image after optimization;

[0099] (7d) According to the saved homography matrix, the image is transformed into a panoramic image, such as Figure 7 shown.

[0100] In summary, in view of the fact that it is difficult to match the feature points of field UAV images and mismatching is likely to occur, the present invention uses the mutual correlation coefficient to detect feature point mismatching; the present invention uses the Delauny algorithm to construct a triangulation network, which can greatly reduce the number of local registrations; the present invention can involve a large number of UAV images; the present invention solves the shortcomings of field images being easily mismatched and too few participating images by taking 132 UAV field images as research objects; feature point mismatching will lead to errors in the calculation of the homography matrix, which has a significant impact on the calculation of the homography matrix of subsequent images. The present invention uses the mutual correlation coefficient to check whether the feature points are mismatched, and eliminates feature point pairs with poor registration effects; the present invention uses the Nelder-Mead algorithm to optimize the registration results, which can reduce the global error to a certain extent.

Claims

1. A field image stitching method based on the Delauny algorithm and the Kruskal maximum spanning tree algorithm, characterized by: The method comprises the following steps in sequence: (1) Obtaining UAV field images and the pose information corresponding to the UAV field images, the pose information of each UAV field image includes internal orientation elements and external orientation elements; the internal orientation elements include the internal parameter matrix, focal length and distortion parameters, and the external orientation elements include X s 、Y s , Z s 、ω、 and γ, where X s 、Y s , Z s is the coordinate of the photography center, ω, γ is the azimuth angle; (2) Reproject each UAV field image according to its interior and exterior orientation elements to obtain the initial projection matrix; (3) The photographic center coordinates X of the drone field image s 、Y s Construct nodes and form a triangulation network using the Delauny algorithm; (4) Through the nodes in the triangulated network, the images corresponding to each two nodes are matched with feature points using the SIFT operator, and then the homography matrix is ​​obtained through the initial projection matrix and projected. If the cross-correlation coefficient after projection is greater than 0.5 or less than -0.5, the cross-correlation coefficient and the corresponding feature point matching results are retained; otherwise, the cross-correlation coefficient and the corresponding feature point matching results are deleted; (5) The mutual correlation coefficients retained in step (4) are used as weights, and the image numbers corresponding to the retained mutual correlation coefficients are used as nodes to form an undirected graph, and the splicing order is generated by the Kruskal maximum spanning tree algorithm and the DFS traversal algorithm; (6) Taking the first node in the splicing order as the reference of the panorama, projecting according to the splicing order and the matching results of the retained feature points to obtain the panorama; (7) Calculate the global registration error of the panorama and optimize the global registration error of the panorama.

2. The method for stitching field images based on the Delauny algorithm and the Kruskal maximum spanning tree algorithm according to claim 1, characterized in that: The step (2) specifically includes the following steps: (2a) First, the UAV field image is dedistorted according to the intrinsic parameter matrix and distortion parameters, as shown in formula (1), and the dedistorted image is saved; Among them, (x, y) is the pixel coordinate, r 2 =x 2 +y 2 , k1, k2, k3 are radial distortion parameters, p1, p2 are tangential distortion parameters, x distortion and y distortion is the distortion amount; (2b) According to the exterior orientation elements of the drone image and the focal length of the drone camera, the WGS84 coordinates of the four corners of the image are obtained, namely the coordinates of the four corner points (x1, y1), (x2, y2), (x3, y3) and (x4, y4). The collinearity principle is used to obtain the collinear coordinates (X1, Y1, Z4) of the four corner points at the average ground elevation. avr )、(X2,Y2,Z avr )、(X3,Y3,Z avr ) and (X4, Y4, Z avr ): Among them, Z avr is the average elevation, f is the focal length, λ is the scaling factor, (2c) Based on the coordinates of the four corner points of the image and the collinear coordinates obtained in step (2b), calculate the corresponding homography matrix and save it as the initial projection matrix: Among them, the initial projection matrix h0 to h7 are elements of the initial projection matrix H.

3. The field image stitching method based on the Delauny algorithm and the Kruskal maximum spanning tree algorithm according to claim 1, characterized in that: The step (3) specifically refers to: the coordinates of each image [X s ,Y s ] as nodes, and generate a triangulation network through the Delauny algorithm. If there are edges between the nodes in the network, it means that the images corresponding to the two nodes may have overlapping parts, and finally save it in the format [node, node].

4. The field image stitching method based on Delauny algorithm and Kruskal maximum spanning tree algorithm according to claim 1, characterized in that: The step (4) specifically includes the following steps: (4a) Using the triangulated network obtained in step (3), the images corresponding to each two nodes i and j are used to extract their respective feature points and feature vectors using the SIFT operator. It is assumed that the projection image i needs to be projected into the coordinate system of the reference image j using the homography matrix; (4b) Perform ratio coarse matching and RANSAC algorithm fine matching on the extracted eigenvectors, and then calculate the homography matrix H i,j ; The obtained homography matrix H i,j , get the homography matrix that projects the projection image into the reference image coordinate system: H i =H i,j H j (4) Among them, H i and H j are the initial projection matrices of images i and j respectively; (4c) Calculate the mutual correlation coefficient of the pixels in the overlapping area, as shown in formula (5). If the mutual correlation coefficient is greater than 0, it indicates that the overlapping similarity is positively correlated, and the closer it is to 1, the higher the correlation. If the mutual correlation coefficient is less than 0, it indicates that the overlapping areas are negatively correlated, and the closer it is to -1, the higher the correlation. If the mutual correlation coefficient is equal to 0, it indicates that the overlapping areas are uncorrelated: Where ncc represents the cross-correlation coefficient, m represents the number of pixels in the overlapping area, and I i,p Indicates the grayscale of the pth pixel in the overlapping area of ​​image i, I j,p represents the grayscale of the p-th pixel in the overlapping area of ​​image j; represents the average grayscale value of pixels in the overlapping area of ​​image i, represents the average grayscale value of pixels in the overlapping area of ​​image j; (4d) If the mutual correlation coefficient is less than 0.5 and greater than -0.5, it indicates that the registration effect is too poor and feature point mismatching is likely to occur. The feature point matching results corresponding to the mutual correlation coefficient less than 0.5 and greater than -0.5 are deleted, and the remaining feature point matching results are saved.

5. The field image stitching method based on Delauny algorithm and Kruskal maximum spanning tree algorithm according to claim 1, characterized in that: The step (5) specifically includes the following steps: (5a) The mutual correlation coefficient retained in step (4) is magnified to 100 times, and the absolute value is taken as the weight E. The image sequence number corresponding to the retained mutual correlation coefficient is used as the node V to form an undirected graph G =<V,E> ; (5b) Generate a maximum spanning tree using the Kruskal maximum spanning tree algorithm, taking the node of the edge with the largest weight as the starting node, and generate the image splicing order using the DFS traversal algorithm.

6. The field image stitching method based on Delauny algorithm and Kruskal maximum spanning tree algorithm according to claim 1, characterized in that: The step (6) specifically includes the following steps: (6a) Taking the first image in the stitching order as the reference of the panorama, the initial projection matrix is ​​used as the homography matrix of the first image; (6b) Match feature points one by one using the SIFT operator in the stitching order, record the homography matrix of the image, and perform image fusion on the overlapping areas. Repeat step (6b) until all images are stitched together to obtain a panoramic image.

7. The field image stitching method based on Delauny algorithm and Kruskal maximum spanning tree algorithm according to claim 1, characterized in that: The step (7) specifically includes the following steps: (7a) After the splicing is completed, the error between the nodes of the undirected graph is calculated: Among them, error i,j represents the local registration error between image i and image j, n is the number of feature point matches between image i and image j; (x i,k ,y i,k ) is the kth feature point of image i, (x j,k ,y j,k ) is the kth feature point of image j, H i (x i,k ,y i,k ) is the kth feature point of image i through the uniqueness matrix H i Transform to the coordinates in the panorama, H j (x j,k ,y j,k ) is the uniqueness matrix H of the kth feature point of image j j Transform to coordinates in the panorama; (7b) Calculate the global registration error: Among them, error represents the global registration error, where i∈[0,s], j∈[0,t]; (7c) using the Nelder-Mead algorithm to optimize the global registration error and save the optimized homography matrix of each image transformed into the panoramic image; (7d) Transform the image into a panorama according to the saved homography matrix.

Images