Large-breadth satellite-borne remote sensing image matching method and system

Abstract

The invention discloses a large-breadth satellite-borne remote sensing image matching method and a large-breadth satellite-borne remote sensing image matching system. According to the method, satellite-borne remote sensing image matching is realized through four parts of image blocking, local image block matching, mismatching point elimination and matching result output. Wherein the image blocking means that a limited number of local image blocks are extracted from the satellite-borne remote sensing image to reduce the calculation amount in the image matching process; local image block matching refers to matching of homonymy points among image blocks by using a currently existing matching method; mismatching point elimination refers to elimination of mismatching points in a matching result after the preliminary matching points are obtained; the matching result output is to selectively output points in the matching result as required. Based on the method, a matching algorithm in the field of computer vision can be easily applied to remote sensing images.

Description

Technical Field

[0001] The invention belongs to the field of remote sensing image processing and is a matching method for relatively large satellite-borne remote sensing images. Background Art

[0002] At present, all image matching methods are designed for small-sized images (generally with a side length of 300-2000). However, satellite remote sensing images are often larger in size in order to balance width and resolution. For example, the GF7 forward-looking image has a length and width of 31267 pixels and 30996 pixels, respectively. When image matching methods are directly applied to satellite remote sensing images, there are two problems: (1) The image data volume is too large. It takes a lot of computation to read and calculate it at one time, which is impossible for most computers to use. (2) The reliability of the matching algorithm is greatly reduced: Such large-sized remote sensing images will extract hundreds of millions of feature points. As the size increases, the specificity of feature points in a single image decreases, which leads to a significant increase in the probability of mismatching of such feature points and a decrease in the reliability of image matching results. Therefore, it is necessary to design a framework for the matching of satellite remote sensing images so that the current matching methods can be easily applied to satellite remote sensing images. Summary of the invention

[0003] The technical problem solved by the present invention is to provide a matching framework for satellite-borne remote sensing images, so that conventional matching algorithms can be well applied to the matching work of satellite-borne remote sensing images.

[0004] The technical solution to achieve the purpose of the present invention is: a wide-bandwidth satellite-borne remote sensing image matching method, comprising the following steps:

[0005] Step 1, dividing the image into blocks based on the image space, that is, dividing the image into blocks based on the image space coordinates of the original image, and extracting image blocks on the reference image based on geographic information;

[0006] Step 2: Use a matching method to match the proposed image blocks and preliminarily obtain the same-name points between the image blocks;

[0007] Step 3, using a local plus global mismatch point elimination strategy to eliminate mismatch points in the matching results;

[0008] Among them, the local RANSAC mismatch point elimination method based on the affine model and the global mismatch point elimination method based on quadratic polynomial residual fitting;

[0009] Step 4: Selectively output the points in the matching results, that is, only one point is selected for each image block involved in the matching to ensure the uniform distribution of the matching results of the entire scene image.

[0010] Furthermore, in step 1, if the width of the original image is w and the length is h, the number of image blocks in the width direction is n x , n in the high direction y , assuming that the size of the image block involved in the matching is s, the coordinates (x, y) of the sampling point are calculated as follows:

[0011]

[0012] x i，j =S x +i*I x

[0013] y i，j =S y +j*I y

[0014] Among them, (S x , S y ) The coordinates of the starting sampling, i and j are the serial numbers of the horizontal sampling and vertical sampling directions, i≤n x , j≤n y ; After obtaining the above image coordinates (x i，j ,y i，j ) and then convert it into object coordinates (Lon i，j , Lat i，j ), so that the coordinate system is converted to the object coordinate system, and then the sampling block can be obtained with the above object coordinate system as the center and s as the side length.

[0015] Furthermore, the specific implementation of step 2 includes two steps. The first step is to eliminate the angle and resolution differences between the image blocks; the second step is to obtain the same-name points of the image blocks after eliminating the angle and resolution differences through a matching algorithm; specifically, it includes the following sub-steps:

[0016] Step 21, eliminating the resolution difference of image rotation: first, calculate the minimum enclosing rectangle of the intersection area of ​​the two images to obtain the four object-space coordinates A, B, C, and D of the minimum enclosing rectangle, then respectively calculate the image-space coordinates A1, B1, C1, and D1 corresponding to the four object-space coordinates of the minimum enclosing rectangle on the original image and the four image-space coordinates A2, B2, C2, and D2 corresponding to the four object-space coordinates of the minimum enclosing rectangle on the reference image, transform the image with the center of the image block as the origin, and subtract the average value of the four corner points of the corresponding image from the above coordinates to obtain the transformation matrix H from the original image block to the reference image block with the center of the image block as the origin; in the subsequent matching process, the original image block is transformed using the transformation matrix H to obtain an image block with the same resolution and angle as the reference image block;

[0017] Step 22: After obtaining image blocks with consistent angles and resolutions, use a matching method to match the image blocks.

[0018] Furthermore, the matching method is divided into feature matching and template matching; wherein, template matching includes template extraction, template similarity calculation, and the most similar template center and the center of the image to be matched are a pair of matching points; feature matching includes feature point extraction, feature point description, and feature point association, wherein feature point association is judged based on the similarity between feature point descriptors, and the two most similar points are a pair of matching points.

[0019] Furthermore, the specific implementation method of removing the local RANSAC mismatched points based on the affine model is as follows:

[0020] ① Randomly select points to solve the affine transformation matrix

[0021] Randomly select three pairs of points with the same name from the matching result point sets X and Y in step 3, and solve the affine transformation matrix H based on the least squares method. The solution method of the transformation matrix is ​​as follows:

[0022] First, match the point pair (x i ,y i ) and (u i , v i ) is expressed as homogeneous coordinates and expressed in matrix form as shown below:

[0023]

[0024] Where X and Y represent the local matching point sets in the original image and the reference image, respectively, and n represents the number of matching pairs;

[0025] Then according to the formula H=Y T X(X T X) -1 Calculate the least squares solution and obtain the transformation matrix H;

[0026] ②Predict the matching point location

[0027] Use the calculated affine transformation matrix H to map all matching points X to obtain the predicted matching point position X′;

[0028] ③ Calculate the number of inner points

[0029] Calculate the Euclidean distance E between the predicted matching point position X′ and the actual matching point X. If E is less than the threshold T, it is considered an inlier. If E is greater than the threshold T, it is considered an outlier. Count the total number of inliers contained in X.

[0030] ④ Repeat steps ① to ③, and take the set of data with the largest number of internal points as the best data set. The point pairs contained in the data set are the locally correctly matched points with the same name.

[0031] Furthermore, the specific implementation of the global mismatch point elimination method based on quadratic polynomial residual fitting is as follows:

[0032] The fitting formula of the quadratic polynomial is as follows:

[0033] A=(X T X) -1 X T L

[0034] Where A is the conversion matrix from the reference image matching points to the original image, X is the matrix composed of the reference image matching points, and L is the matrix composed of the corresponding points of the original image. The construction method of each matrix is ​​as follows:

[0035] A=[a0 a1 a2 ... a 10 a 11 ] T

[0036]

[0037] L=[x′1 y′1 ... x′ k y′ k ] T

[0038] Where (x k ,y k ) is the kth feature point on the reference image; (x′ k , y′ k ) is the corresponding point of the kth point of the reference image matched in the original image, and a0-a11 represent the conversion coefficients constituting the conversion matrix A;

[0039] After obtaining the transformation matrix A, the Euclidean distance between the coordinates of the same-name points obtained based on the transformation matrix and the matched coordinates of the same-name points is calculated for each point. On this basis, the root mean square error of all distances is calculated, and points with distances greater than m times the root mean square error are removed as mismatched points.

[0040] Furthermore, the calculation formula of the root mean square error R is as follows:

[0041]

[0042] Where (x i0 ,y i0 ) is the coordinate of the feature point in the reference image; d i is the Euclidean distance between the coordinates of the same-name point in the original image obtained based on the transformation matrix and the coordinates of the matched same-name point of the i-th feature point in the reference image; (x, y) and (x′, y′) represent the coordinates of the same-name point in the i-th feature point in the reference image obtained based on the transformation matrix and the coordinates of the matched same-name point, respectively;

[0043] Judgement i The relationship between the size of m×R is that if d i ≥mR, the i-th point is considered to be a mismatched point, otherwise it is retained; after all feature points are judged in turn, the mismatched points are eliminated.

[0044] Furthermore, the specific implementation of step 4 is as follows:

[0045] The points are screened one by one for each image block. For a single image block, the center coordinates (C x , C y ), the calculation formula of the center coordinates is as follows:

[0046] Max x =max(x1',x2',x3',...,x n ')

[0047] Min x =min(x1', x2', x3', ..., x n ')

[0048] Max y =max(y1',y2',y3',...,y n ')

[0049] Min y =min(y1',y2',y3',...,y n ')

[0050]

[0051] In the formula, (x' i ,y' i ) is the i-th coordinate of a single image block in the reference image, and n represents the number of matching points in the image block;

[0052] After obtaining the coordinates of the center point, calculate the matching points one by one and (C x ,C y ), and output the point with the smallest Euclidean distance as the final matching point of the image block;

[0053] By performing the above operations on each image block in turn, matching points evenly distributed in the image can be obtained.

[0054] The present invention also provides a wide-bandwidth satellite-borne remote sensing image matching system, comprising the following modules:

[0055] The image segmentation module is used to segment the image based on the image space, that is, to segment the image based on the image space coordinates of the original image, and to extract image blocks on the reference image based on geographic information;

[0056] The image block matching module is used to match the proposed image blocks using a matching method and preliminarily obtain the same-name points between the image blocks;

[0057] The wrong matching point elimination module is used to eliminate the wrong matching points in the matching results by adopting a local plus global wrong matching point elimination strategy;

[0058] Among them, the local RANSAC mismatch point elimination method based on the affine model and the global mismatch point elimination method based on quadratic polynomial residual fitting;

[0059] The point aggregation output module is used to selectively output the points in the matching results, that is, only one point is selected for each image block involved in the matching to ensure the uniform distribution of the matching results of the entire scene image.

[0060] The innovation of the present invention lies in: (1) designing a matching method for remote sensing images, based on which the matching algorithm in the field of computer vision can be easily applied to remote sensing images.

[0061] (2) A method for eliminating mismatched points in remote sensing images is designed, namely, a mismatched point elimination strategy that combines the initial elimination of RANSAC mismatched points based on the affine transformation model of local image blocks with the secondary elimination of mismatched points based on geographic location information. This strategy can effectively improve the reliability of the output results.

[0062] (3) A point aggregation strategy was designed to ensure the uniformity of the distribution of points with the same name in remote sensing image matching while ensuring the matching efficiency of remote sensing images. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 Schematic diagram of a flow chart of an embodiment of the present invention.

[0064] Figure 2 The following is a schematic diagram of points with the same name matched in GF7 satellite image data based on an embodiment of the present invention. DETAILED DESCRIPTION

[0065] The specific implementation modes of the present invention are described in detail below with reference to the accompanying drawings and embodiments.

[0066] Satellite-borne remote sensing images are relatively wide, with image length and width often reaching tens of thousands. However, all current matching algorithms are designed for small image blocks, and there are few matching algorithms for such large amounts of data. Therefore, how to apply these algorithms to satellite-borne remote sensing images is a problem worth solving. In response to the above problems, the present invention constructs a set of matching methods and systems for satellite-borne remote sensing images. The framework achieves the matching of satellite-borne remote sensing images through four parts: image segmentation, local image block matching, mismatch point elimination, and matching result output.

[0067] See also Figure 1 The process of the embodiment can be divided into the following steps, and the specific method, formula and process for implementing each step are as follows:

[0068] Step 1, divide the image into blocks based on the image space. This step refers to dividing the image into blocks based on the image space coordinates of the original image, and extracting image blocks on the reference image based on geographic information. If the width of the original image is w and the length is h, the number of image blocks in the width direction is n x , ny in the high direction, assuming that the size of the image block involved in the matching is s, the coordinates (x, y) of the sampling point are calculated as follows:

[0069]

[0070] x i，j =S x +i*I x

[0071] y i，j =S y +j*I y

[0072] After obtaining the above image coordinates (x i，j ,y i，j ) and then convert it into object coordinates (Lon i，j , Lat i，j ), so that the coordinate system is converted to the object coordinate system, and then the sampling block is obtained with this series of coordinates as the center and s as the side length, which lays the foundation for the subsequent extraction of the image blocks involved in the matching, where (S x , S y ) The coordinates of the starting sampling, usually i and j are the numbers of horizontal sampling and vertical sampling directions (i≤n x , j≤n y )).

[0073] Step 2: Use the matching method to match the proposed image blocks and preliminarily obtain the same-name points between the image blocks. This part is mainly divided into two steps. The first step is to eliminate the angle and resolution differences between the image blocks; the second step is to obtain the same-name points of the image blocks after eliminating the angle and resolution differences through the matching algorithm. The specific steps are as follows:

[0074] Step 21, image rotation resolution difference elimination, first, calculate the minimum bounding rectangle of the intersection area of ​​the two images to obtain the four object coordinates A, B, C, D of the minimum bounding rectangle, then calculate the image coordinates A1, B1, C1, D1 corresponding to the four object coordinates of the minimum bounding rectangle on the original image and the four image coordinates A2, B2, C2, D2 corresponding to the reference image, respectively. In order to make the calculated transformation matrix transform the image with the center of the image block as the origin, we subtract the average value of the four corner points of the corresponding image from the above coordinates, so as to obtain the transformation matrix H from the original image block to the reference image block with the center of the image block as the origin. In the subsequent matching process, the original image block is transformed using the transformation matrix H to obtain an image block with the same resolution and angle as the reference image block.

[0075] Step 22, after obtaining the image blocks with the same angle and resolution, the image blocks are matched by a matching method. The matching methods can be mainly divided into feature matching and template matching. Among them, template matching includes template extraction, template similarity calculation, and the most similar template center and the center of the image to be matched are a pair of matching points; feature matching includes feature point extraction, feature point description, and feature point association, wherein the feature point association is judged based on the similarity between the feature point descriptors, and the two most similar points are a pair of matching points. At this point, the same-name points between the local image blocks are matched.

[0076] Step 3, local to global mismatch point elimination. After the above method, the same-name points between different image blocks are obtained, but there are inevitably mismatch points between these same-name points. In view of this, a local + global mismatch point elimination strategy is adopted to eliminate the mismatch points in the matching results. The elimination method is as follows.

[0077] Step 31, local RANSAC mismatch point elimination based on affine model. Although the satellite remote sensing image has a wide width, the transformation relationship between the local parts of the two images is relatively simple. The RANSAC method based on affine model can be used to eliminate the mismatch points, which specifically includes the following steps:

[0078] ① Randomly select points to solve the affine transformation matrix

[0079] Randomly select three pairs of points with the same name from the matching result point sets X and Y in 3.2, and solve the affine transformation matrix H based on the least squares method. The solution method of the transformation matrix is ​​as follows:

[0080] First, match the point pair (x i ,y i ) and (u i , v i ) is expressed as homogeneous coordinates and expressed in matrix form as shown below

[0081]

[0082] Where X and Y represent the local matching point sets in the original image and the reference image respectively, and n represents the number of matching pairs.

[0083] Then according to the formula H=Y T X(X T X) -1 Calculate the least squares solution and obtain the transformation matrix H.

[0084] ②Predict the matching point location

[0085] Use the calculated affine transformation matrix H to map all matching points X to obtain the predicted matching point position X′.

[0086] ③ Calculate the number of inner points

[0087] Calculate the Euclidean distance E between the predicted matching point position X′ and the actual matching point X. If E is less than the threshold T (set to 3 in this algorithm), it is considered an internal point, and if E is greater than the threshold T, it is considered an external point. Count the total number of internal points contained in X.

[0088] ④ Repeat steps ① to ③, and take the data set with the largest number of internal points as the best data set. The point pairs contained in this data set are the locally correctly matched points with the same name. In order to take into account the robustness and efficiency of the algorithm, this framework sets the number of loops to 5000.

[0089] Through the above operations, the incorrect matching points of the local image blocks can be eliminated.

[0090] Step 32: Global mismatch point elimination method based on quadratic polynomial residual fitting

[0091] After the local mismatch points are eliminated, the accuracy of the matching results will be greatly improved. However, there will still be some significant mismatch points. In addition, due to the influence of the RANSAC random sampling mechanism, when the internal point rate in the local image block is low, the screened points will be all wrong matching points. In order to minimize this situation, we further use a global mismatch point elimination method based on quadratic polynomial residual fitting to eliminate significant outliers to improve the accuracy of the final matching results. Specifically, due to the wide width of satellite remote sensing images, it is difficult to fit the matching results through a simple affine model. Therefore, we use quadratic polynomials to fit the transformation relationship between images, and eliminate outliers in the matching results based on the fitting results. The fitting formula of the quadratic polynomial is as follows:

[0092] A=(X T X) -1 X T L

[0093] Where A is the conversion matrix from the reference image matching points to the original image, X is the matrix composed of the reference image matching points, and L is the matrix composed of the corresponding points of the original image. The construction method of each matrix is ​​as follows:

[0094] A=[a0 a1 a2 ... a 10 a 11 ] T

[0095]

[0096] L=[x′1 y′1 ... x′ k y′ k ] T

[0097] Where (x k ,y k ) is the kth feature point on the reference image; (x′ k , y′ k ) is the corresponding point of the kth point in the reference image matched in the original image, and a0-a11 represent the conversion coefficients constituting the conversion matrix A.

[0098] After obtaining the transformation matrix A, the Euclidean distance between the coordinates of the same-name points obtained based on the transformation matrix and the coordinates of the matched same-name points is calculated for each point. On this basis, the root mean square error of all distances is calculated, and points with a distance greater than 3 times (can also be set according to specific needs) of the root mean square error are removed as mismatched points. The calculation formula of the root mean square error R is as follows:

[0099]

[0100] Where (x i0 ,y i0 ) is the coordinate of the feature point in the reference image; d i is the Euclidean distance between the coordinates of the homonymous point in the original image obtained based on the transformation matrix and the matched coordinates of the homonymous point of the i-th feature point in the reference image; (x, y) and (x′, y′) represent the coordinates of the homonymous point in the i-th feature point in the reference image obtained based on the transformation matrix and the matched coordinates of the homonymous point, respectively.

[0101] Judgement i The relationship between d and 3×R is i If ≥3R, the i-th point is considered as a mismatch point, otherwise it is retained. After all feature points are judged in sequence, the removal of mismatch points is completed.

[0102] Step 4: Selectively output the points in the matching results. Matching point screening mainly refers to selecting only one point from each image block involved in the matching to ensure the uniform distribution of the matching results of the entire scene image. The specific operation is: screen the points one by one in the image block. Taking a single image block as an example, we first calculate the center coordinates of the image block (C x , C y ), the calculation formula of the center coordinates is as follows:

[0103] Max x =max(x1',x2',x3',...,x n ')

[0104] Min x =min(x1', x2', x3', ..., x n ')

[0105] Max y =max(y1',y2',y3',...,y n ')

[0106] Min y =min(y1',y2',y3',...,y n ')

[0107]

[0108] In the formula, (x' i , y' i ) is the i-th coordinate of a single image block in the reference image, and n represents the number of matching points in the image block.

[0109] After obtaining the center point coordinates, we calculate the matching points one by one and (C x , C y) and output the point with the smallest Euclidean distance as the final matching point of the image block. Perform the above operations on each image block in turn to obtain matching points evenly distributed in the image. The experiment verified that the GF7 satellite's front and back image matching has a total of 164 feature points, which takes 5.393 seconds. The matching results are shown in the attached figure. Figure 2 As shown in the figure, the matching points are evenly distributed in the entire image, which can meet the needs of most satellite remote sensing image matching.

[0110] The advantages of the present invention are: the satellite-borne remote sensing images are sampled and divided into blocks to reduce the amount of calculation in the matching process, and the local plus global mismatch point elimination method is used to effectively eliminate the mismatch points in the satellite-borne remote sensing images with large distortion. Finally, the matching point selection strategy is used to effectively ensure the uniformity of the image matching point distribution.

[0111] On the other hand, an embodiment of the present invention further provides a wide-bandwidth satellite-borne remote sensing image matching system, comprising the following modules:

[0112] The image segmentation module is used to segment the image based on the image space, that is, to segment the image based on the image space coordinates of the original image, and to extract image blocks on the reference image based on geographic information;

[0113] The image block matching module is used to match the proposed image blocks using a matching method and preliminarily obtain the same-name points between the image blocks;

[0114] The wrong matching point elimination module is used to eliminate the wrong matching points in the matching results by adopting a local plus global wrong matching point elimination strategy;

[0115] Among them, the local RANSAC mismatch point elimination method based on the affine model and the global mismatch point elimination method based on quadratic polynomial residual fitting;

[0116] The point aggregation output module is used to selectively output the points in the matching results, that is, only one point is selected for each image block involved in the matching to ensure the uniform distribution of the matching results of the entire scene image.

[0117] The specific implementation method of each module is the same as each step and will not be described in detail in the present invention.

[0118] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A wideband satellite-borne remote sensing image matching method, characterized in that: The following steps are involved: Step 1, dividing the image into blocks based on the image space, that is, dividing the image into blocks based on the image space coordinates of the original image, and extracting image blocks on the reference image based on geographic information; Step 2: Use a matching method to match the proposed image blocks and preliminarily obtain the same-name points between the image blocks; Step 3, using a local plus global mismatch point elimination strategy to eliminate mismatch points in the matching results; Among them, the local RANSAC mismatch point elimination method based on the affine model and the global mismatch point elimination method based on quadratic polynomial residual fitting; Step 4: Selectively output the points in the matching results, that is, only one point is selected for each image block involved in the matching to ensure the uniform distribution of the matching results of the entire scene image.

2. The wide-bandwidth satellite-borne remote sensing image matching method according to claim 1, characterized in that: In step 1, if the width of the original image is w and the length is h, the number of image blocks in the width direction is n x , n in the high direction y , assuming that the size of the image block involved in the matching is s, the coordinates (x, y) of the sampling point are calculated as follows: x i,j =S x +i*I x y i,j =S y +j*I y Among them, (S x ,S y ) The coordinates of the starting sampling, i and j are the serial numbers of the horizontal sampling and vertical sampling directions, i≤n x ,j≤n y ; After obtaining the above image coordinates (x i,j ,y i,j ) and then convert it into object coordinates (Lon i,j , Lat i,j ), so that the coordinate system is converted to the object coordinate system, and then the sampling block can be obtained with the above object coordinate system as the center and s as the side length.

3. The wide-bandwidth satellite-borne remote sensing image matching method according to claim 1, characterized in that: The specific implementation of step 2 includes two steps. The first step is to eliminate the angle and resolution differences between image blocks. The second step is to obtain the same-name points of the image blocks after eliminating the angle and resolution differences through a matching algorithm. The specific sub-steps are as follows: Step 21, eliminating the resolution difference of image rotation: first, calculate the minimum enclosing rectangle of the intersection area of ​​the two images to obtain the four object-space coordinates A, B, C, and D of the minimum enclosing rectangle, then respectively calculate the image-space coordinates A1, B1, C1, and D1 corresponding to the four object-space coordinates of the minimum enclosing rectangle on the original image and the four image-space coordinates A2, B2, C2, and D2 corresponding to the four object-space coordinates of the minimum enclosing rectangle on the reference image, transform the image with the center of the image block as the origin, and subtract the average value of the four corner points of the corresponding image from the above coordinates to obtain the transformation matrix H from the original image block to the reference image block with the center of the image block as the origin; in the subsequent matching process, the original image block is transformed using the transformation matrix H to obtain an image block with the same resolution and angle as the reference image block; Step 22: After obtaining image blocks with consistent angles and resolutions, use a matching method to match the image blocks.

4. The wide-bandwidth satellite-borne remote sensing image matching method according to claim 3, characterized in that: The matching methods are divided into feature matching and template matching; among them, template matching includes template extraction, template similarity calculation, and the most similar template center and the center of the image to be matched are a pair of matching points; feature matching includes feature point extraction, feature point description, and feature point association, among which feature point association is judged based on the similarity between feature point descriptors, and the two most similar points are a pair of matching points.

5. The wide-bandwidth satellite-borne remote sensing image matching method according to claim 1, characterized in that: The specific implementation method of removing RANSAC mismatched points based on the local affine model is as follows: ① Randomly select points to solve the affine transformation matrix Randomly select three pairs of points with the same name from the matching result point sets X and Y in step 3, and solve the affine transformation matrix H based on the least squares method. The solution method of the transformation matrix is ​​as follows: First, match the point pair (x i ,y i ) and (u i ,v i ) is expressed as homogeneous coordinates and expressed in matrix form as shown below: Where X and Y represent the local matching point sets in the original image and the reference image, respectively, and n represents the number of matching pairs; Then according to the formula H=Y T X(X T X) -1 Calculate the least squares solution and obtain the transformation matrix H; ②Predict the matching point location Use the calculated affine transformation matrix H to map all matching points X to obtain the predicted matching point position X′; ③ Calculate the number of inner points Calculate the Euclidean distance E between the predicted matching point position X′ and the actual matching point X. If E is less than the threshold T, it is considered an inlier. If E is greater than the threshold T, it is considered an outlier. Count the total number of inliers contained in X. ④ Repeat steps ① to ③, and take the set of data with the largest number of internal points as the best data set. The point pairs contained in the data set are the locally correctly matched points with the same name.

6. The wide-bandwidth satellite-borne remote sensing image matching method according to claim 1, characterized in that: The specific implementation of the global mismatch point elimination method based on quadratic polynomial residual fitting is as follows: The fitting formula of the quadratic polynomial is as follows: A=(X T X) -1 X T L Where A is the conversion matrix from the reference image matching points to the original image, X is the matrix composed of the reference image matching points, and L is the matrix composed of the corresponding points of the original image. The construction method of each matrix is ​​as follows: A=[a0a1a2…a 10 a 11 ] T L=[x′1y′1…x′ k the k ] T Where (x k ,y k ) is the kth feature point on the reference image; (x′ k ,y′ k ) is the corresponding point of the kth point in the reference image matched in the original image, a0-a 11 represents the conversion coefficients constituting the conversion matrix A; After obtaining the transformation matrix A, the Euclidean distance between the coordinates of the same-name points obtained based on the transformation matrix and the matched coordinates of the same-name points is calculated for each point. On this basis, the root mean square error of all distances is calculated, and points with distances greater than m times the root mean square error are removed as mismatched points.

7. The wide-bandwidth satellite-borne remote sensing image matching method according to claim 6, characterized in that: The calculation formula of the root mean square error R is as follows: Where (x i0 ,y i0 ) is the coordinate of the feature point in the reference image; d i is the Euclidean distance between the coordinates of the same-name point in the original image obtained based on the transformation matrix and the coordinates of the matched same-name point of the i-th feature point in the reference image; (x, y) and (x′, y′) represent the coordinates of the same-name point in the i-th feature point in the reference image obtained based on the transformation matrix and the coordinates of the matched same-name point, respectively; Judgement i The relationship between the size of m×R is that if d i ≥mR, the i-th point is considered to be a mismatched point, otherwise it is retained; after all feature points are judged in turn, the mismatched points are eliminated.

8. The wide-bandwidth satellite-borne remote sensing image matching method according to claim 1, characterized in that: The specific implementation of step 4 is as follows: The points are screened one by one for each image block. For a single image block, the center coordinates (C x ,C y ), the calculation formula of the center coordinates is as follows: Max x =max(x1’,x2’,x3’,…,x n ’) My x =min(x1',x2',x3',…,x n ') Max y =max(y1’,y2’,y3’,…,y n ’) <h2 style=";text-align:left;direction:ltr">Min<h2 style=";text-align:left;direction:ltr"> y <h2 style=";text-align:left;direction:ltr"> = min(y1',y2',y3',…,y<h2 style=";text-align:left;direction:ltr"> n <h2 style=";text-align:left;direction:ltr"> ') In the formula, (x' i ,y' i ) is the i-th coordinate of a single image block in the reference image, and n represents the number of matching points in the image block; After obtaining the coordinates of the center point, calculate the matching points one by one and (C x ,C y ), and output the point with the smallest Euclidean distance as the final matching point of the image block; By performing the above operations on each image block in turn, matching points evenly distributed in the image can be obtained.

9. A wideband satellite-borne remote sensing image matching system, characterized in that: Includes the following modules: The image segmentation module is used to segment the image based on the image space, that is, to segment the image based on the image space coordinates of the original image, and to extract image blocks on the reference image based on geographic information; The image block matching module is used to match the proposed image blocks using a matching method and preliminarily obtain the same-name points between the image blocks; The wrong matching point elimination module is used to eliminate the wrong matching points in the matching results by adopting a local plus global wrong matching point elimination strategy; Among them, the local RANSAC mismatch point elimination method based on the affine model and the global mismatch point elimination method based on quadratic polynomial residual fitting; The point aggregation output module is used to selectively output the points in the matching results, that is, only one point is selected for each image block involved in the matching to ensure the uniform distribution of the matching results of the entire scene image.

Images