Large-scene SAR image matching method based on deep learning

Abstract

The invention provides a large-scene SAR image matching method based on deep learning. The method comprises the following steps: S1, acquiring a first estimated displacement field and a first matched SAR image of a first reference SAR image and a first to-be-processed SAR image by adopting a matching model; s2, acquiring a first common area, a first effective reference SAR image and a first to-be-matched SAR image; s3, acquiring a second estimated displacement field and a second matched SAR image; s4, acquiring a second common area, a second effective reference SAR image and a second to-be-matched SAR image; s5, dividing the SAR image into a plurality of SAR image blocks, inputting the SAR image blocks into a matching model, and obtaining a third estimated displacement field and a third matched SAR image; s6, constructing a verification SAR image block pair; obtaining a fourth displacement field and a verification displacement field of the SAR image block pair in the matching model; s7, determining an overlapping center region and a displacement deviation of each group of SAR image block pairs and the verification SAR image block; extracting homonymy point pairs; and S8, fitting a coordinate transformation model of the third matching SAR image and the first reference SAR image to obtain a final matching SAR image.

Description

Technical Field

[0001] The present invention relates to the technical field of image matching, and in particular to a large-scene SAR image matching method based on deep learning. Background Art

[0002] Synthetic Aperture Radar (SAR) is an active microwave remote sensing imaging radar with the advantages of all-weather, all-day, and high resolution. It can be installed on satellites, aircraft, ships and other platforms, and used in military, agriculture, disaster monitoring, resource exploration and other fields. Therefore, it is receiving more and more attention.

[0003] SAR image matching is the process of matching two SAR images of the same scene acquired at different times or angles. Existing SAR image matching can be roughly divided into two categories: grayscale matching and feature matching. Grayscale-based SAR image matching is mainly based on the similarity of the pixel grayscale between the real-time image to be matched and the pre-stored reference image, and finds the geometric transformation with the maximum similarity between the two images. This type of method is relatively simple and easy to implement, but it is significantly affected by noise. Feature-based SAR image matching methods extract obvious and stable features from the real-time image to be matched and the reference image, then match the features and calculate the geometric transformation parameters. This type of method has strong adaptability to the grayscale changes of the image, but due to the coherent speckle noise in the SAR image, it is easy to make the extraction of feature points wrong, resulting in wrong matching. In addition, there is also the problem of difficulty in matching in large-scene SAR images due to the large differences in offset and change rules in different regions. In order to solve the above problems, it is urgent to develop a method that can efficiently and accurately achieve SAR image matching in large scenes. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a large scene SAR image matching method based on deep learning, the method comprising:

[0005] S1 sequentially analyzes the first reference SAR image I according to the preset ratio B and the first SAR image to be processed I A After downsampling, the trained matching model is input to obtain the first estimated displacement field F1 and the first matching SAR image T 1 (I A );

[0006] S2 obtains the first reference SAR image I B and the first matching SAR image T 1 (I A ) of the first common area; and the first reference SAR image I in the first common area BThe area is taken as the first valid reference SAR image The first matching SAR image T in the first common area 1 (I A ) is taken as the first SAR image to be matched

[0007] S3 respectively performs the first effective reference SAR image according to the preset ratio and the first SAR image to be matched After downsampling, it is input into the trained matching model to obtain the second estimated displacement field F2 and the second matching SAR image T 2 (I A );

[0008] S4 Acquire the first reference SAR image I B and the second matching SAR image T 2 (I A ) of the second common area; and the first reference SAR image I in the second common area B The area is used as the second effective reference SAR image Belongs to the second matching image T 2 (I A ) is used as the second SAR image to be matched

[0009] S5 sequentially converts the second valid reference SAR image and the second SAR image to be matched The image is divided into a plurality of SAR image blocks of a preset size to obtain a plurality of second SAR image block pairs; each group of second SAR image block pairs is sequentially input into the trained matching model to obtain an average displacement field And the mean displacement field Perform upsampling and padding processing to obtain the third estimated displacement field F3 and the third matching SAR image T 3 (I A );

[0010] S6 respectively takes the first reference SAR image I B and the third matching SAR image T 3 (I A ) is divided into a plurality of SAR image blocks of preset sizes to obtain a plurality of groups of third SAR image block pairs; a verification SAR image block corresponding to each group of third SAR image block pairs is constructed; each group of third SAR image block pairs and the corresponding verification SAR image block are sequentially input into the trained matching model respectively, to obtain a fourth displacement field for each group of third SAR image block pairs and verify the displacement field

[0011] S7: Determine the overlapping central area of ​​each group of third SAR image block pairs and the corresponding verification SAR image block, and obtain a fourth displacement field of each group of third SAR image block pairs. and verify the displacement field The displacement deviation in the overlapping central area; and extracting the same-name point pairs of each group of third SAR image block pairs according to the displacement deviation to obtain the same-name point pair set;

[0012] S8 Fitting the third matching SAR image T based on the set of points with the same name 3 (I A ) and the first reference SAR image I B The coordinate transformation model between the third matching SAR image T 3 (I A ) to perform coordinate transformation and interpolation resampling to obtain the final matching SAR image T(I A ).

[0013] Specifically, the matching model is a GLU-Net model, which includes an H-Net network and an L-Net network; the L-Net network includes a feature extraction layer, a global correlation layer, and a displacement field estimator, and the H-Net network includes a feature extraction layer, a local correlation layer, and a refinement network; wherein the feature extraction layer is a pre-trained VGG-16 network; the refinement network is a feedforward convolutional neural network, which includes multiple dilated convolutional layers; the displacement field estimator includes multiple convolutional layers, residual blocks, and a DenseNet connection structure.

[0014] Specifically, the L-Net network is used to process the downsampled reference SAR image and the SAR image to be matched, and uses the global correlation layer to realize the geometric transformation between the SAR images at low resolution, and outputs the displacement estimation field; the H-Net network is used to refine the displacement estimation field output by the L-Net network through the local correlation layer and the refinement network, and obtain the displacement field between the reference SAR image and the SAR image to be matched at the original resolution.

[0015] A two-way matching training strategy is adopted to train the matching model based on a sample pair set; each sample pair in the sample set includes a reference SAR sample image X B and the SAR sample image X to be matched A ; The two-way matching training strategy includes forward matching training and reverse matching training; during the training process, a comprehensive loss function L is introduced to calculate the gap between the predicted displacement field and the actual displacement field until the loss function tends to be stable, thereby generating a trained matching model.

[0016] Specifically, the training method of the matching model includes:

[0017] A two-way matching training strategy is adopted to train the matching model based on a sample pair set; each sample pair in the sample pair set includes a reference SAR sample image X B and the SAR sample image X to be matched A ; The two-way matching training strategy includes forward matching training and reverse matching training; during the training process, a comprehensive loss function L is introduced to calculate the gap between the predicted displacement field and the actual displacement field until the loss function tends to be stable, thereby generating a trained matching model.

[0018] The forward matching training includes: in each group of sample pairs, using the reference SAR sample image X B As a benchmark, the reference SAR sample image X B and the SAR sample image X to be matched A Input the matching model to obtain the first predicted displacement field Based on the first predicted displacement field Get the first matching SAR sample image T(X A );

[0019] The reverse matching training includes: using the SAR sample image X to be matched A As a benchmark, the SAR sample image to be matched and the reference SAR sample image X B Input into the matching model and obtain the second predicted displacement field Based on the second predicted displacement field Get the second matching SAR sample image

[0020] Then use the second matching SAR sample image As the reference, the second matching SAR sample image and the first matching SAR sample image T(X A ) and the input matching model, the third predicted displacement field is obtained Based on the third predicted displacement field Get T(T(X A )).

[0021] In this embodiment of the present invention, the comprehensive loss function of the matching model is:

[0022]

[0023] in, is the loss function in forward matching training; is the loss function in reverse matching training; It is expressed as:

[0024]

[0025] in, is the first predicted displacement field in the forward matching training; is the first real displacement field in the forward matching training; H×W is the reference SAR sample image X B size; It is expressed as:

[0026]

[0027] in, A second predicted displacement field in reverse matching training; is the second true displacement field in forward matching training.

[0028] Specifically, step S7 further includes:

[0029] S71 determining an overlapping central area between each group of third SAR image block pairs and a corresponding verification SAR image block;

[0030] S72 Obtain the fourth displacement field of each group of third SAR image block pairs and verify the displacement field The displacement deviation in the overlapping center area;

[0031] S73 selects a plurality of groups of target image block pairs from the plurality of groups of third SAR image block pairs according to the displacement deviation;

[0032] S74 Assume that the coordinates of the center point of each group of target image blocks on the first reference SAR image IB are (x ci ,y ci ), then the center point corresponds to the first SAR image to be matched I A The coordinates of the points on the same name are (x ci +Δx i ,y ci +Δy i );

[0033] in

[0034] S75 will be in the first reference SAR image I B The center point on the first SAR image to be matched I A The points with the same name on the are regarded as a pair of points with the same name;

[0035] S76 repeats the above steps S71-S75 until the same-name point pairs of each group of target block pairs are obtained to obtain a set of same-name point pairs.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] (1) The present invention uses the GLU-Net matching model, which includes an L-Net network and an H-Net network, makes full use of global and local related feature information, gradually refines the displacement field estimation, and finally realizes accurate estimation of the displacement field at different scales;

[0038] (2) The present invention first estimates the displacement field between the SAR image to be matched and the reference SAR image using the reference SAR image as a benchmark; then estimates the displacement field between the SAR sample image to be matched and the reference SAR sample image using the SAR image to be matched as a benchmark, and adopts a two-way matching training strategy to train the matching model, thereby ensuring the geometric consistency and reversibility in the image matching process and improving the accuracy of the model;

[0039] (3) The present invention solves the problem of matching difficulty caused by large differences in offsets and change patterns in different regions in SAR images of large scenes, and provides a method for efficiently and accurately realizing SAR image matching in large scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0041] Figure 1 It is a technical flow chart of a large-scene SAR image matching method based on deep learning in an embodiment of the present invention;

[0042] Figure 2 A technical schematic diagram of constructing a verification SAR image block corresponding to each group of third SAR image block pairs in an embodiment of the present invention;

[0043] Figure 3 A schematic diagram of the structure of a matching model in an embodiment of the present invention;

[0044] Figure 4 It is a technical flow chart of the two-way matching training strategy in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field belong to the scope of protection of the present invention.

[0046] See also Figure 1 , Figure 1 This is a technical flow chart of a large-scene SAR image matching method based on deep learning in an embodiment of the present invention. The method includes:

[0047] S1 sequentially analyzes the first reference SAR image I according to the preset ratio B and the first SAR image to be processed I A After downsampling, the trained matching model is input to obtain the first estimated displacement field F1 and the first matching SAR image T 1 (I A ).

[0048] In an embodiment of the present invention, the first reference SAR image I B and the first SAR image to be matched I A The size is H×W, proportional It is downsampled to a size of h×w, and the first downsampled displacement field is obtained in the trained network model. Then the first estimated displacement field F1 corresponding to the image size can be expressed as:

[0049]

[0050] Where mean(F'1) means taking the mean of F'1 in the first two dimensions, Up() means upsampling operation, and the height and width after sampling are H×W. Represents the adjustment coefficient of the displacement field value range, so that the displacement range corresponds to the size of the original image. A , we can get the first matching SAR image

[0051] S2 obtains the first reference SAR image I B and the first matching SAR image T 1 (I A ) of the first common area; and the first reference SAR image I in the first common area B The area is taken as the first valid reference SAR image The first matching SAR image T in the first common area 1 (I A ) is taken as the first SAR image to be matched

[0052] In an embodiment of the present invention, according to the first SAR image to be matched Determine the position and size of the effective area, and thereby intercept the first reference SAR image I B and the first matching SAR image T 1 (I A ). Assume that the image obtained is and Where 0≤y1<y2≤H and 0≤x1<x2≤W, the image size is H valid ×W valid .

[0053] S3 respectively performs the first effective reference SAR image according to the preset ratio and the first SAR image to be matched After downsampling, it is input into the trained matching model to obtain the second estimated displacement field F2 and the second matching SAR image T 2 (I A ).

[0054] In an embodiment of the present invention, the first valid reference SAR image and the first SAR image to be matched According to the preset ratio The resulting image (size is denoted as h valid × valid ) is input into the trained matching model to obtain the second down-sampled displacement field Transform it to get the first transformation displacement field Then ΔF valid It can be expressed as:

[0055]

[0056] According to the location and size of the effective area, F valid The edges are copied and filled to obtain Then the second estimated displacement field F2 after the second step adjustment is:

[0057] F2=ΔF+F1

[0058] Based on F2 and I A Get the second matching SAR image

[0059] S4 Acquire the first reference SAR image I B and the second matching SAR image T 2 (I A ) of the second common area; and the first reference SAR image I in the second common area B The area is used as the second effective reference SAR image Belongs to the second matching image T 2 (I A ) is used as the second SAR image to be matched

[0060] S5 sequentially converts the second valid reference SAR image and the second SAR image to be matched The image is divided into a plurality of SAR image blocks of a preset size to obtain a plurality of second SAR image block pairs; each group of second SAR image block pairs is sequentially input into the trained matching model to obtain an average displacement field And the mean displacement field Perform upsampling and padding processing to obtain the third estimated displacement field F3 and the third matching SAR image T 3 (I A ).

[0061] In the embodiment of the present invention, the above steps all involve downsampling the reference image and the image to be matched to a certain size before inputting into the network model, so the image resolution is lost, resulting in a relatively rough displacement field. The present invention adopts block processing to avoid downsampling the input image, so as to make a more refined correction to the existing displacement field. First, according to the second matching SAR image T 2 (I A ) determine the position and size of the effective area, and intercept the first reference SAR image I B and the second matching SAR image T 2 (I A ) is the second common area, whose size is denoted by H valid ×W valid Secondly, the second valid reference SAR image and the second SAR image to be matched Divide into non-overlapping second SAR image block pairs of size 2000×2000. Assume that the total number of second SAR image block pairs is The i-th group of image block pairs is input into the trained network model to obtain the block displacement field of each group of second SAR image block pairs: and its mean where i=1,2,..., All the means obtained Arranged as the mean displacement field Upsample it to get the second transformed displacement field

[0062]

[0063] According to the location and size of the effective area, ΔF" valid The edges are copied and filled to obtain Then the third estimated displacement field for:

[0064] F3=F2+ΔF'

[0065] Based on F3 and I A, get three matching SAR images T 3 (I A ).

[0066] S6 respectively takes the first reference SAR image I B and the third matching SAR image T 3 (I A ) is divided into a plurality of SAR image blocks of preset sizes to obtain a plurality of groups of third SAR image block pairs; a verification SAR image block corresponding to each group of third SAR image block pairs is constructed; each group of third SAR image block pairs and the corresponding verification SAR image block are sequentially input into the trained matching model respectively, to obtain a fourth displacement field for each group of third SAR image block pairs and verify the displacement field

[0067] See also Figure 2 , Figure 2 Schematic diagram of a technique for constructing a verification SAR image block corresponding to each group of third SAR image block pairs in an embodiment of the present invention; in an embodiment of the present invention, according to the third matching SAR image T 3 (I A ) determine the position and size of the effective area, and intercept the first reference SAR image I B and the third matching SAR image T 3 (I A ) is the third common area, whose size is denoted by H valid ×W valid Secondly, the first reference SAR image I B and the third matching SAR image T 3 (I A ) is divided into non-overlapping SAR image blocks of size 512×512 to form multiple groups of third SAR image block pairs; each group of third SAR image block pairs includes a first reference SAR image block and a third matching SAR image block. Based on each group of third SAR image block pairs, its corresponding verification SAR image block is constructed; the verification SAR image block pair is obtained by translating the reference SAR image block by 64 pixels along the x direction. Each group of third SAR image block pairs and the corresponding verification SAR image block pair are input into the trained matching model in turn to obtain the fourth displacement field of each group of third SAR image block pairs and verify the displacement field

[0068] S7: Determine the overlapping central area of ​​each group of third SAR image block pairs and the corresponding verification SAR image block, and obtain a fourth displacement field of each group of third SAR image block pairs. and verify the displacement field The displacement deviation in the overlapping central area is extracted according to the displacement deviation, and the same-name point pairs of each group of third SAR image blocks are obtained to obtain the same-name point pair set.

[0069] In this embodiment of the present invention, step S7 further includes:

[0070] S71 determining an overlapping central area between each group of third SAR image block pairs and a corresponding verification SAR image block;

[0071] S72 Obtain the fourth displacement field of each group of third SAR image block pairs and verify the displacement field The displacement deviation in the overlapping center area;

[0072] S73 selects a plurality of groups of target image block pairs from the plurality of groups of third SAR image block pairs according to the displacement deviation.

[0073] In the embodiment of the present invention, the displacement field of the central area of ​​the image block is usually more accurate. and The displacement deviation in the overlapping central area is used to determine the matching of each set of third SAR image block pairs. Let the size of the central area be 256×256. In the overlapping central area, the calculation formula for the proportion of pixels whose displacement deviation meets the requirements is:

[0074]

[0075] in The value of th is set to 1, Ω represents the overlapping center area, and N Ω is the total number of pixels in the central overlapping area. ratio <ratio i When the third SAR image block pair meets the matching requirements, th ratio Set to 0.95.

[0076] S74 assumes that each group of target image blocks is centered on the first reference SAR image I B The coordinates of the upper center point are (x ci ,y ci ), then the center point corresponds to the first SAR image to be matched I A The coordinates of the points on the same name are (x ci +Δx i ,y ci +Δy i );

[0077] in

[0078] S75 will be in the first reference SAR image I BThe center point on the first SAR image to be matched I A The points with the same name on the are regarded as a pair of points with the same name;

[0079] S76 repeats the above steps S71-S75 until the same-name point pairs of each group of target block pairs are obtained to obtain a set of same-name point pairs.

[0080] S8 Fitting the third matching SAR image T based on the set of points with the same name 3 (I A ) and the first reference SAR image I B The coordinate transformation model between the third matching SAR image T 3 (I A ) to perform coordinate transformation and interpolation resampling to obtain the final matching SAR image T(I A ).

[0081] In the embodiment of the present invention, the first reference SAR image and the first SAR image to be processed Formulate the image matching problem as a dense displacement field (It can also be expressed as the displacement field in the x direction Displacement field in the y direction ) is estimated, then the final matching SAR image T(I A ) can be expressed as:

[0082] T(I A )(x,y)=I A (x+d x (x,y),y+d y (x,y)

[0083] Where (x, y) represents the final matching SAR image T(I A ), (x+d x (x,y),y+d y (x, y) represents the corresponding first SAR image to be processed I A Coordinates of points of the same name on .

[0084] See also Figure 3 , Figure 3It is a structural schematic diagram of the matching model in the embodiment of the present invention; in the embodiment of the present invention, the matching model is a GLU-Net model, which includes an H-Net network and an L-Net network; the L-Net network includes a feature extraction layer, a global correlation layer, and a displacement field estimator, and the H-Net network includes a feature extraction layer, a local correlation layer and a refinement network; wherein the feature extraction layer is a pre-trained VGG-16 network; the refinement network is a feedforward convolutional neural network, which includes multiple dilated convolutional layers; the displacement field estimator includes multiple convolutional layers, a residual block and a DenseNet connection structure.

[0085] In an embodiment of the present invention, the L-Net network is used to process the downsampled reference SAR image and the SAR image to be matched, and uses the global correlation layer to realize the geometric transformation between the SAR images at low resolution, and outputs the displacement estimation field; the H-Net network is used to refine the displacement estimation field output by the L-Net network through the local correlation layer and the refinement network, and obtain the displacement field between the reference SAR image and the SAR image to be matched at the original resolution.

[0086] See also Figure 4 , Figure 4 : is a technical flow chart of the two-way matching training strategy in an embodiment of the present invention; in an embodiment of the present invention, the specific method for training the matching model includes:

[0087] A two-way matching training strategy is adopted to train the matching model based on a sample pair set; each sample pair in the sample pair set includes a reference SAR sample image X B and the SAR sample image X to be matched A ; The two-way matching training strategy includes forward matching training and reverse matching training; during the training process, a comprehensive loss function L is introduced to calculate the gap between the predicted displacement field and the actual displacement field until the loss function tends to be stable, thereby generating a trained matching model.

[0088] The forward matching training includes: in each group of sample pairs, using the reference SAR sample image X B As a benchmark, the reference SAR sample image X B and the SAR sample image X to be matched A Input the matching model to obtain the first predicted displacement field Based on the first predicted displacement field Get the first matching SAR sample image T(X A );

[0089] The reverse matching training includes: using the SAR sample image X to be matched A As a benchmark, the SAR sample image to be matched and the reference SAR sample image X BInput into the matching model and obtain the second predicted displacement field Based on the second predicted displacement field Get the second matching SAR sample image

[0090] Then use the second matching SAR sample image As the benchmark, the second matching SAR sample image and the first matching SAR sample image T(X A ) and the input matching model, the third predicted displacement field is obtained Based on the third predicted displacement field Get T(T(X A )).

[0091] In this embodiment of the present invention, the comprehensive loss function of the matching model is:

[0092]

[0093] in, is the loss function in forward matching training; is the loss function in reverse matching training; It is expressed as:

[0094]

[0095] in, is the first predicted displacement field in the forward matching training; is the first real displacement field in the forward matching training; H×W is the reference SAR sample image X B size; It is expressed as:

[0096]

[0097] in, A second predicted displacement field in reverse matching training; is the second true displacement field in forward matching training.

[0098] Those skilled in the art may make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A large scene SAR image matching method based on deep learning, characterized in that: The method comprises: S1 sequentially analyzes the first reference SAR image I according to the preset ratio B and the first SAR image to be processed I A After downsampling, the trained matching model is input to obtain the first estimated displacement field F1 and the first matching SAR image T 1 (I A ); S2 obtains the first reference SAR image I B and the first matching SAR image T 1 (I A ) of the first common area; and the first reference SAR image I in the first common area B The area is taken as the first valid reference SAR image The first matching SAR image T in the first common area 1 (I A ) is taken as the first SAR image to be matched S3 respectively performs the first effective reference SAR image according to the preset ratio and the first SAR image to be matched After downsampling, it is input into the trained matching model to obtain the second estimated displacement field F2 and the second matching SAR image T 2 (I A ); S4 Acquire the first reference SAR image I B and the second matching SAR image T 2 (I A ) of the second common area; and the first reference SAR image I in the second common area B The area is used as the second effective reference SAR image Belongs to the second matching image T 2 (I A ) is used as the second SAR image to be matched S5 sequentially converts the second valid reference SAR image and the second SAR image to be matched The image is divided into a plurality of SAR image blocks of a preset size to obtain a plurality of second SAR image block pairs; each group of second SAR image block pairs is sequentially input into the trained matching model to obtain an average displacement field And the mean displacement field Perform upsampling and padding processing to obtain the third estimated displacement field F3 and the third matching SAR image T 3 (I A ); S6 respectively takes the first reference SAR image I B and the third matching SAR image T 3 (I A ) is divided into a plurality of SAR image blocks of preset sizes to obtain a plurality of groups of third SAR image block pairs; a verification SAR image block corresponding to each group of third SAR image block pairs is constructed; each group of third SAR image block pairs and the corresponding verification SAR image block are sequentially input into the trained matching model respectively, to obtain a fourth displacement field for each group of third SAR image block pairs and verify the displacement field S7: Determine the overlapping central area of ​​each group of third SAR image block pairs and the corresponding verification SAR image block, and obtain a fourth displacement field of each group of third SAR image block pairs. and verify the displacement field The displacement deviation in the overlapping central area; and extracting the same-name point pairs of each group of third SAR image block pairs according to the displacement deviation to obtain the same-name point pair set; S8 Fitting the third matching SAR image T based on the set of points with the same name 3 (I A ) and the first reference SAR image I B The coordinate transformation model between the third matching SAR image T 3 (I A ) to perform coordinate transformation and interpolation resampling to obtain the final matching SAR image T(I A ).

2. The method according to claim 1, characterized in that The matching model is a GLU-Net model, which includes an H-Net network and an L-Net network; the L-Net network includes a feature extraction layer, a global correlation layer, and a displacement field estimator, and the H-Net network includes a feature extraction layer, a local correlation layer, and a refinement network; wherein the feature extraction layer is a pre-trained VGG-16 network; the refinement network is a feedforward convolutional neural network, which includes multiple dilated convolutional layers; the displacement field estimator includes multiple convolutional layers, a residual block, and a DenseNet connection structure.

3. The method according to claim 2, characterized in that The L-Net network is used to process the downsampled reference SAR image and the SAR image to be matched, and uses the global correlation layer to realize the geometric transformation between the SAR images at low resolution, and outputs the displacement estimation field; the H-Net network is used to refine the displacement estimation field output by the L-Net network through the local correlation layer and the refinement network, and obtain the displacement field between the reference SAR image and the SAR image to be matched at the original resolution.

4. The method according to claim 3, characterized in that The specific method for training the matching model includes: A two-way matching training strategy is adopted to train the matching model based on a sample pair set; each sample pair in the sample pair set includes a reference SAR sample image X B and the SAR sample image X to be matched A ; The two-way matching training strategy includes forward matching training and reverse matching training; during the training process, a comprehensive loss function L is introduced to calculate the gap between the predicted displacement field and the actual displacement field until the loss function tends to be stable, thereby generating a trained matching model.

5. The method according to claim 4, characterized in that The forward matching training includes: in each group of sample pairs, using the reference SAR sample image X B As a benchmark, the reference SAR sample image X B and the SAR sample image X to be matched A Input the matching model to obtain the first predicted displacement field Based on the first predicted displacement field Get the first matching SAR sample image T(X A ); The reverse matching training includes: using the SAR sample image X to be matched A As a benchmark, the SAR sample image to be matched and the reference SAR sample image X B Input into the matching model and obtain the second predicted displacement field Based on the second predicted displacement field Get the second matching SAR sample image Then use the second matching SAR sample image As the benchmark, the second matching SAR sample image and the first matching SAR sample image T(X A ) and the input matching model, the third predicted displacement field is obtained Based on the third predicted displacement field Get T(T(X A )).

6. The method according to claim 5, characterized in that The comprehensive loss function of the matching model is: in, is the loss function in forward matching training; is the loss function in reverse matching training; It is expressed as: in, is the first predicted displacement field in the forward matching training; is the first real displacement field in the forward matching training; H×W is the reference SAR sample image X B size; It is expressed as: in, A second predicted displacement field in reverse matching training; is the second true displacement field in forward matching training.

7. The method according to claim 1, characterized in that Step S7 also includes: S71 determining an overlapping central area between each group of third SAR image block pairs and a corresponding verification SAR image block; S72 obtains the fourth displacement field of each group of third SAR image block pairs and verify the displacement field The displacement deviation in the overlapping center area; S73 selects a plurality of groups of target image block pairs from the plurality of groups of third SAR image block pairs according to the displacement deviation; S74 Assume that the coordinates of the center point of each group of target image blocks on the first reference SAR image IB are (x ci ,y ci ), then the center point corresponds to the first SAR image I to be matched A The coordinates of the points on the same name are (x ci +Δx i ,y ci +Δy i ); in S75 will be in the first reference SAR image I B The center point on the first SAR image to be matched I A The points with the same name on the are regarded as a pair of points with the same name; S76 repeats the above steps S71-S75 until the same-name point pairs of each group of target block pairs are obtained to obtain a set of same-name point pairs.

Images