SAR image adaptive registration method based on GLU-Net-Evalue

Abstract

The invention relates to the field of image processing, in particular to an SAR (synthetic aperture radar) image adaptive registration method based on GLU-Net-Evaluate. According to the technical problems, when SAR images are registered, a mature visible light image registration model is applied to SAR image registration, due to the fact that the characteristic and distribution differences between the visible light images and the SAR images are large, the effect is often poor, and due to the fact that the data scarcity problem in the aspect of SAR image registration is serious, the SAR images cannot be registered. Therefore, not only is the training and verification of the algorithm impacted, but also the problems of over-fitting and poor generalization performance are possibly caused. According to the technical scheme, the SAR image self-adaptive registration method based on the GLU-Net-Evalue comprises the steps that a final prediction optical flow field is generated through an optical flow estimation network, so that image registration is completed; according to the SAR image registration method, the GLU-Net network is trained by setting the transfer learning technology, then training is carried out on scarce SAR data, the evaluation network is added, the evaluation network is used for guiding training on the SAR, the limitation of an existing method can be overcome, and the performance of SAR image registration is improved.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a SAR image adaptive registration method based on GLU-Net-Evaluate. Background Art

[0002] Traditional SAR image registration research can be divided into two major methods: feature matching-based and region-based methods. The core idea of ​​the region-based method is to define a similarity measurement criterion, use the intensity information of the image to calculate the similarity between the current window and the template at each position in the image to be registered, and estimate the geometric transformation model by optimizing the similarity between the reference image and the image to be registered. These methods usually perform image registration in a specific transform domain (such as Fourier domain, wavelet domain, etc.). For example, frequency domain matching uses Fourier transform to convert the image to the frequency domain and then perform registration; while wavelet transform uses wavelet transform to extract multi-scale features of the image for registration. Commonly used similarity measurement criteria include mutual information, cross-accumulated residual entropy, normalized mutual information, etc. However, the region-based SAR image registration method is weak in processing the geometric differences and radiation differences of SAR images, has high computational complexity, and usually requires a long computing time, which limits its wide application. In contrast, deep learning-based registration methods have obvious development prospects. These methods take advantage of deep learning and can learn from large-scale samples. SAR images can be used to extract richer deep semantic information; compared with traditional methods, SAR registration methods based on deep learning can capture more complex image features and establish more sophisticated image space mapping relationships; through the feature representation of deep learning, the registration method can mine more abstract and discriminative features from SAR images, making the common features between SAR images more prominent and achieving more accurate image matching; this deep learning-based method not only performs well in cases where samples are limited or scenes are complex, but also can show excellent performance in large-scale data and diversified scenes; therefore, deep learning methods have broad application potential, especially for specific application scenarios, rather than a general solution; when registering SAR images, the mature visible light image registration model is applied to SAR image registration. Due to the large differences in characteristics and distribution between visible light images and SAR images, the effect is often not good, and due to the serious problem of data scarcity in SAR image registration, it not only has an impact on the training and verification of the algorithm, but also may lead to overfitting and poor generalization performance. Summary of the invention

[0003] In order to overcome the problem that the mature visible light image registration model is applied to SAR image registration when registering SAR images, the effect is often poor due to the large differences in characteristics and distribution between visible light images and SAR images. In addition, due to the serious data scarcity problem in SAR image registration, it not only affects the training and verification of the algorithm, but also may lead to overfitting and poor generalization performance.

[0004] The technical solution of the present invention is: a SAR image adaptive registration method based on GLU-Net-Evaluate, comprising the following steps:

[0005] S11: performing data preprocessing on the input SAR image data, converting different SAR image data into standardized image data;

[0006] S12: extracting features from the preprocessed image data using a feature network;

[0007] S13: performing feature matching on the SAR image data after feature extraction, and finding the same feature points or feature areas in multiple SAR image data;

[0008] S14: Use the optical flow estimation network to generate the final predicted optical flow field, and use the predicted optical flow field to complete image registration.

[0009] Preferably, the GLU-Net network is used to obtain the predicted optical flow field between the two images, and the image to be registered is generated based on the predicted optical flow field to complete the image registration task; this process includes data preprocessing, feature extraction, feature matching, optical flow estimation and calculation of optical flow loss function.

[0010] Preferably, when data preprocessing is performed on the input SAR image data to convert different SAR image data into standardized image data, the input image includes an active image Is and a target image It.

[0011] Preferably, when performing data preprocessing on the input SAR image data to convert different SAR image data into standardized image data, the following steps are included:

[0012] S21: First, the source image Is and the target image It are adjusted to appropriate resolutions; wherein the formula for adjusting the resolution is:

[0013] Is'=adjustment resolution (Is), It'=adjustment resolution (It),

[0014] Wherein, Is' is the source image after adjusting the resolution, It' is the target image after adjusting the resolution;

[0015] S22: performing standardization processing on the source image Is' and the target image It' after adjusting the resolution; wherein the principle formula of the standardization processing is:

[0016] Is” = Standardization (Is'), It” = Standardization (It');

[0017] Among them, Is” is the source image after the resolution is adjusted and standardized, and It” is the target image after the resolution is adjusted and standardized.

[0018] Preferably, when using a feature network to extract features from preprocessed image data, the feature network used is a VGG network; wherein the VGG network uses a 3×3 convolution kernel and increases the network depth to extract the source image and target image after adjusting the resolution and standardizing the processing.

[0019] Preferably, the VGG network uses a VGG-16 model to extract features from the preprocessed image data; wherein the form of the VGG network is:

[0020] Conv3×3→ReLU→MaxPool2×2;

[0021] Among them, → indicates that these layers are connected sequentially, Conv is the convolution layer, ReLU is the activation function, and MaxPool is the maximum pooling layer.

[0022] Preferably, the structure of the VGG network is: Input (preprocessed image) → (Conv3×3→ReLU)×2→MaxPool→(Conv3×3→ReLU)×2→MaxPool→......→FC→Output (feature output); the VGG network effectively captures the multi-scale information of the image through multi-layer convolution and pooling; these features are crucial for the image registration task because they provide sufficient information to find the correspondence between different images; the purpose of the feature extraction layer is to capture local and global information in the image, which is crucial for subsequent image registration tasks; after feature extraction, a high-dimensional feature space that can describe the image content and structure is obtained, which greatly improves the model's ability to perform accurate registration in complex scenes.

[0023] Preferably, when feature matching is performed on the SAR image data after feature extraction and the same feature points or feature regions are found in multiple SAR image data, feature matching is performed in a global-local feature matching manner.

[0024] Preferably, when performing feature matching, global feature matching is mainly responsible for overall large-scale motion, while local feature matching is responsible for subtle local motion. The entire GLU-Net feature M matching is a process based on a collaborative mechanism, which means that global and local feature matching are performed simultaneously so that more comprehensive matching information can be obtained at different scales and fields of view.

[0025] Preferably, when performing feature matching, a special correlation layer Correlation is used to perform feature matching, and the correlation layer Correlation is used to calculate the correlation between different features; wherein the correlation layer Correlation calculates the inner product of a feature of one image and a feature of another image, and generates a correlation graph, thereby achieving feature matching.

[0026] Preferably, the correlation layer Correlation uses the rectangular image patch image1patch of the source image to perform a convolution operation on the image2patch of the target image when calculating the correlation between different features.

[0027] Preferably, the specific calculation process of Correlation is based on the convolution operation in CNN, but the convolution operation in CNN is deformed. Compared with the use of a specific convolution kernel for convolution in CNN, the rectangular image block patch (image1patch) of the first image is used to perform a convolution operation on the patch (image2patch) of another image, so the operation does not contain training parameters; for a rectangular image block (patch) of image1 (width w, height h, number of channels c) with a length k (the length and width of the rectangular block is 2k+1) extending up and down with x1 as the center, a Correlation calculation between it and a rectangular image block (patch) in image2 (w, h, c) with a length k extending up and down with x2 as the center can be expressed as:

[0028] c(x1,x2)=∑ o∈[-k，k]×[-k，k] (f1(x1+o),f2(x2+o));

[0029] In detail, the point x1 in image1(w, h, c) means that a pixel point centered on x1 is selected from the first image image1, and the "rectangular image patch with a length k extending up and down" here means that instead of considering the entire image, a small area around the point x1 is considered, which is a square patch with a side length of 2k+1; this size is based on "k", which is a parameter that defines the radius of the patch; therefore, if k is 1, then the patch size will be 3x3; a correlation between it and the rectangular image patch with a length k extending up and down and centered on x2 in image2(w, h, c) means that, The corresponding patch centered at x2 is selected in the second image image2; what needs to be calculated is the correlation between the patch centered at x1 in image1 and the patch centered at x2 in image2; this calculation will compare the two patches and produce a numerical value indicating the similarity or degree of match between them; this correlation is calculated by comparing the patch of image1 with all possible patches in image2; in order to find the best matching area, in the optical flow prediction, this best matching area will tell us where x1 has moved in image2, so that the movement of pixel x1 between two consecutive frames, that is, the optical flow, can be calculated.

[0030] Preferably, when the final predicted optical flow field is generated by using the optical flow estimation network and the image registration is completed by using the predicted optical flow field, the following steps are included:

[0031] S31: first, extract features from the input target image It and the source image Is using a feature extractor at different resolutions, wherein the different resolutions include L1, L2, L3 and L4, and the resolution levels of L1, L2, L3 and L4 increase in sequence;

[0032] S32: Then at resolution L1, the network calculates the global correlation between the source feature FL1 and the target feature FL1, and uses normalization and ReLU activation functions to process the obtained global correlation to strengthen the determined matches and reduce the weights of uncertain matches;

[0033] S33: Global correlation is mapped to decoder M top Converted into a two-dimensional dense correspondence map m1, which is then converted into a displacement field Get the rough optical flow from the source image to the target image;

[0034] S34: The calculated local correlation module is used on different resolution layers L2, L3 and L4 in turn to further refine the optical flow. At the same time, the optical flow decoder uses the local correlation and the optical flow sampled in the previous layer to calculate the residual optical flow, and then the residual optical flow is added to the optical flow of the previous layer to obtain the complete optical flow of the current layer.

[0035] Preferably, when the optical flow estimation network is used to generate the final predicted optical flow field and the predicted optical flow field is used to complete the image registration, the evaluation network is used to update and optimize the parameters of the optical flow estimation network, wherein the evaluation network uses LSTM to generate a scene for evaluation and inputs it into the loss calculation function to guide the overall parameter update by integrating historical information. The workflow of the evaluation network is:

[0036] S41: receiving optical flow estimation data of the image as network input;

[0037] S42: Network input data is processed through fully connected layers, batch normalization layers, and residual connections to achieve feature extraction and enhancement;

[0038] S43: The extracted features are regularized by the Dropout layer to obtain a preliminary evaluation field;

[0039] S44: LSTM layer processes the preliminary evaluation field;

[0040] S45: The LSTM output passes through two fully connected layers to generate the evaluation field.

[0041] Preferably, the key advantage of introducing the evaluation field is that it can accelerate the convergence speed of the model during the training process, especially in the face of data scarcity. Through the introduction of the evaluation field, the model can learn and adapt to the changing data distribution more quickly, thereby improving the training efficiency and registration performance of the model.

[0042] Preferably, the LSTM model designed in the present algorithm has an important function, which can continuously capture and update the evaluation field features of each pair of images in historical iterations to help output the evaluation field of the pair of images at the current moment; in the order of training iterations, each pair of images sequentially passes its historical evaluation field to the LSTM model to integrate information at each historical moment; this helps to increase the information provided by each pair of images in the current training round, thereby filling the gap in information shortage when data is scarce; in this way, the LSTM model outputs the evaluation field of the current training round, effectively guiding the optical flow estimation network to better complete the SAR image registration task.

[0043] Beneficial effects of the present invention:

[0044] 1. In the existing technology, when registering SAR images, the mature visible light image registration model is applied to SAR image registration. However, due to the large differences in characteristics and distribution between visible light images and SAR images, the effect is often poor. In addition, due to the serious data scarcity problem in SAR image registration, it not only has an impact on the training and verification of the algorithm, but also may cause overfitting and poor generalization performance. The SAR image registration method trains the GLU-Net network by setting the transfer learning technology, and then trains it on scarce SAR data, and adds an evaluation network to guide its training on SAR, which can improve the performance of SAR image registration by using the evaluation network to overcome the limitations of the existing methods;

[0045] 2. The key advantage of introducing the evaluation field is that it can accelerate the convergence speed of the model during the training process. Especially in the case of data scarcity, the model can learn and adapt to the changing data distribution more quickly through the introduction of the evaluation field, thereby improving the training efficiency and registration performance of the model;

[0046] 3. The LSTM model designed in this algorithm has an important function. It can continuously capture and update the evaluation field features of each pair of images in historical iterations to help output the evaluation field of the pair of images at the current moment. In the order of training iterations, each pair of images passes its historical evaluation field to the LSTM model in turn to integrate the information of each historical moment. This helps to increase the information provided by each pair of images in the current training round, thereby filling the gap in information shortage when data is scarce. In this way, the LSTM model outputs the evaluation field of the current training round, effectively guiding the optical flow estimation network to better complete the SAR image registration task. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 What is shown is a flow chart of the SAR image adaptive registration method based on GLU-Net-Evaluate of the present invention;

[0048] Figure 2 What is shown is a specific flow chart of the SAR image adaptive registration method based on GLU-Net-Evaluate of the present invention;

[0049] Figure 3 What is shown is a model schematic diagram of the evaluation network in the SAR image adaptive registration method based on GLU-Net-Evaluate of the present invention;

[0050] Figure 4 What is presented is a schematic diagram of the workflow of the evaluation network in the GLU-Net-Evaluate-based SAR image adaptive registration method of the present invention. DETAILED DESCRIPTION

[0051] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0052] See also Figure 1-2 The present invention provides an embodiment: a SAR image adaptive registration method based on GLU-Net-Evaluate, including the following steps:

[0053] S11: performing data preprocessing on the input SAR image data, converting different SAR image data into standardized image data;

[0054] S12: extracting features from the preprocessed image data using a feature network;

[0055] S13: performing feature matching on the SAR image data after feature extraction, and finding the same feature points or feature areas in multiple SAR image data;

[0056] S14: Use the optical flow estimation network to generate the final predicted optical flow field, and use the predicted optical flow field to complete image registration.

[0057] Preferably, the GLU-Net network is used to obtain the predicted optical flow field between the two images, and the image to be registered is generated based on the predicted optical flow field to complete the image registration task; this process includes data preprocessing, feature extraction, feature matching, optical flow estimation and calculation of optical flow loss function.

[0058] Preferably, when data preprocessing is performed on the input SAR image data to convert different SAR image data into standardized image data, the input image includes an active image Is and a target image It.

[0059] Preferably, when performing data preprocessing on the input SAR image data to convert different SAR image data into standardized image data, the following steps are included:

[0060] S21: First, the source image Is and the target image It are adjusted to appropriate resolutions; wherein the formula for adjusting the resolution is:

[0061] Is'=adjustment resolution (Is), It'=adjustment resolution (It),

[0062] Wherein, Is' is the source image after adjusting the resolution, It' is the target image after adjusting the resolution;

[0063] S22: performing standardization processing on the source image Is' and the target image It' after adjusting the resolution; wherein the principle formula of the standardization processing is:

[0064] Is” = Standardization (Is'), It” = Standardization (It');

[0065] Among them, Is” is the source image after the resolution is adjusted and standardized, and It” is the target image after the resolution is adjusted and standardized.

[0066] Preferably, when using a feature network to extract features from preprocessed image data, the feature network used is a VGG network; wherein the VGG network uses a 3×3 convolution kernel and increases the network depth to extract the source image and target image after adjusting the resolution and standardizing the processing.

[0067] Preferably, the VGG network uses a VGG-16 model to extract features from the preprocessed image data; wherein the form of the VGG network is:

[0068] Conv3×3→ReLU→MaxPool2×2;

[0069] Among them, → indicates that these layers are connected sequentially, Conv is the convolution layer, ReLU is the activation function, and MaxPool is the maximum pooling layer.

[0070] Preferably, the structure of the VGG network is: Input (preprocessed image) → (Conv3×3→ReLU)×2 → MaxPool → (Conv3×3→ReLU)×2 → MaxPool → … → FC → Output (feature output); the VGG network effectively captures the multi-scale information of the image through multiple layers of convolution and pooling; these features are crucial for the image registration task because they provide sufficient information to find the correspondence between different images; the purpose of the feature extraction layer is to capture local and global information in the image, which is crucial for subsequent image registration tasks; after feature extraction, a high-dimensional feature space that can describe the image content and structure is obtained, which greatly improves the model's ability to perform accurate registration in complex scenes.

[0071] Preferably, when feature matching is performed on the SAR image data after feature extraction and the same feature points or feature regions are found in multiple SAR image data, feature matching is performed in a global-local feature matching manner.

[0072] Preferably, when performing feature matching, global feature matching is mainly responsible for overall large-scale motion, while local feature matching is responsible for subtle local motion. The entire GLU-Net feature M matching is a process based on a collaborative mechanism, which means that global and local feature matching are performed simultaneously so that more comprehensive matching information can be obtained at different scales and fields of view.

[0073] Preferably, when performing feature matching, a special correlation layer Correlation is used to perform feature matching, and the correlation layer Corre1ation is used to calculate the correlation between different features; wherein the correlation layer Correlation calculates the inner product of a feature of one image and a feature of another image, and generates a correlation graph, thereby achieving feature matching.

[0074] Preferably, when calculating the correlation between different features, the correlation layer Correlation uses the rectangular image patch image1patch of the source image to perform a convolution operation on the image2patch of the target image.

[0075] Preferably, the specific calculation process of Correlation is based on the convolution operation in CNN, but the convolution operation in CNN is deformed. Compared with the use of a specific convolution kernel for convolution in CNN, the rectangular image block patch (image1patch) of the first image is used to perform a convolution operation on the patch (image2patch) of another image, so the operation does not contain training parameters; for a rectangular image block (patch) of image1 (width w, height h, number of channels c) with a length k (the length and width of the rectangular block is 2k+1) extending up and down with x1 as the center, a Correlation calculation between it and a rectangular image block (patch) in image2 (w, h, c) with a length k extending up and down with x2 as the center can be expressed as:

[0076] c(x1,x2)=∑ o∈[-k，k]×[-k，k] (f1(x1+o),f2(x2+o));

[0077] In detail, the point x1 in image1(w, h, c) means that a pixel point centered on x1 is selected from the first image image1, and the "rectangular image patch with a length k extending up and down" here means that instead of considering the entire image, a small area around the point x1 is considered. This area is a square patch with a side length of 2k+1; this size is based on "k", which is a parameter that defines the radius of the patch; therefore, if k is 1, then the patch size will be 3x3; a correlation between it and the rectangular image patch with a length k extending up and down and centered on x2 in image2(w, h, c) means that, The corresponding patch centered at x2 is selected in the second image image2; what needs to be calculated is the correlation between the patch centered at x1 in image1 and the patch centered at x2 in image2; this calculation will compare the two patches and produce a numerical value indicating the similarity or degree of match between them; this correlation is calculated by comparing the patch of image1 with all possible patches in image2; in order to find the best matching area, in the optical flow prediction, this best matching area will tell us where x1 has moved in image2, so that the movement of pixel x1 between two consecutive frames, that is, the optical flow, can be calculated.

[0078] See also Figure 3-4 In this embodiment, when the final predicted optical flow field is generated by using the optical flow estimation network and the image registration is completed by using the predicted optical flow field, the following steps are included:

[0079] S31: first, extract features from the input target image It and the source image Is using a feature extractor at different resolutions, wherein the different resolutions include L1, L2, L3 and L4, and the resolution levels of L1, L2, L3 and L4 increase in sequence;

[0080] S32: Then at resolution L1, the network calculates the global correlation between the source feature FL1 and the target feature FL1, and uses normalization and ReLU activation functions to process the obtained global correlation to strengthen the determined matches and reduce the weights of uncertain matches;

[0081] S33: Global correlation is mapped to decoder M top is converted into a two-dimensional dense correspondence map m1, which is then converted into a displacement map Get the rough optical flow from the source image to the target image;

[0082] S34: The calculated local correlation module is used on different resolution layers L2, L3 and L4 in turn to further refine the optical flow. At the same time, the optical flow decoder uses the local correlation and the optical flow sampled in the previous layer to calculate the residual optical flow, and then the residual optical flow is added to the optical flow of the previous layer to obtain the complete optical flow of the current layer.

[0083] Preferably, when the optical flow estimation network is used to generate the final predicted optical flow field and the predicted optical flow field is used to complete the image registration, the evaluation network is used to update and optimize the parameters of the optical flow estimation network, wherein the evaluation network uses LSTM to generate a scene for evaluation and inputs it into the loss calculation function to guide the overall parameter update by integrating historical information. The workflow of the evaluation network is:

[0084] S41: receiving optical flow estimation data of the image as network input;

[0085] S42: Network input data is processed through fully connected layers, batch normalization layers, and residual connections to achieve feature extraction and enhancement;

[0086] S43: The extracted features are regularized by the Dropout layer to obtain a preliminary evaluation field;

[0087] S44: LSTM layer processes the preliminary evaluation field;

[0088] S45: The LSTM output passes through two fully connected layers to generate the evaluation field.

[0089] Preferably, the key advantage of introducing the evaluation field is that it can accelerate the convergence speed of the model during the training process, especially in the face of data scarcity. Through the introduction of the evaluation field, the model can learn and adapt to the changing data distribution more quickly, thereby improving the training efficiency and registration performance of the model.

[0090] Preferably, the LSTM model designed in the present algorithm has an important function, which can continuously capture and update the evaluation field features of each pair of images in historical iterations to help output the evaluation field of the pair of images at the current moment; in the order of training iterations, each pair of images sequentially passes its historical evaluation field to the LSTM model to integrate information at each historical moment; this helps to increase the information provided by each pair of images in the current training round, thereby filling the gap in information shortage when data is scarce; in this way, the LSTM model outputs the evaluation field of the current training round, effectively guiding the optical flow estimation network to better complete the SAR image registration task.

[0091] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge scope of those skilled in the art without departing from the spirit of the present invention.

Claims

1. A SAR image adaptive registration method based on GLU-Net-Evaluate; characterized by: The following steps are included: S11: performing data preprocessing on the input SAR image data, converting different SAR image data into standardized image data; S12: extracting features from the preprocessed image data using a feature network; S13: performing feature matching on the SAR image data after feature extraction, and finding the same feature points or feature areas in multiple SAR image data; S14: Use the optical flow estimation network to generate the final predicted optical flow field, and use the predicted optical flow field to complete image registration.

2. The SAR image adaptive registration method based on GLU-Net-Evaluate according to claim 1, characterized in that: When data preprocessing is performed on the input SAR image data to convert different SAR image data into standardized image data, the input image includes an active image Is and a target image It.

3. The SAR image adaptive registration method based on GLU-Net-Evaluate according to claim 2, characterized in that: When the input SAR image data is preprocessed to convert different SAR image data into standardized image data, the following steps are included: S21: First, the source image Is and the target image It are adjusted to appropriate resolutions; wherein the formula for adjusting the resolution is: Is'=adjustment resolution (Is), It'=adjustment resolution (It), Wherein, Is' is the source image after adjusting the resolution, It' is the target image after adjusting the resolution; S22: performing standardization processing on the source image Is' and the target image It' after adjusting the resolution; wherein the principle formula of the standardization processing is: Is” = Standardization (Is'), It” = Standardization (It'); Among them, Is” is the source image after the resolution is adjusted and standardized, and It” is the target image after the resolution is adjusted and standardized.

4. The SAR image adaptive registration method based on GLU-Net-Evaluate according to claim 3, characterized in that: When using the feature network to extract features from the preprocessed image data, the feature network used is the VGG network; wherein the VGG network uses a 3×3 convolution kernel and increases the network depth to extract the source image and target image after adjusting the resolution and standardizing the processing.

5. The SAR image adaptive registration method based on GLU-Net-Evaluate according to claim 4, characterized in that The VGG network uses the VGG-16 model to extract features from the preprocessed image data; the form of the VGG network is: Conv3×3→ReLU→MaxPool2×2; Among them, → indicates that these layers are connected sequentially, Conv is the convolution layer, ReLU is the activation function, and MaxPool is the maximum pooling layer.

6. The SAR image adaptive registration method based on GLU-Net-Evaluate according to claim 5, characterized in that: When feature matching is performed on the SAR image data after feature extraction, when the same feature points or feature areas are found in multiple SAR image data, feature matching is performed using a global-local feature matching method.

7. The SAR image adaptive registration method based on GLU-Net-Evaluate according to claim 6, characterized in that: When performing feature matching, a special correlation layer Correlation is used to perform feature matching. The correlation layer Correlation is used to calculate the correlation between different features. Among them, the correlation layer Correlation calculates the inner product of the feature of one image and the feature of another image, and generates a correlation graph to achieve feature matching.

8. The SAR image adaptive registration method based on GLU-Net-Evaluate according to claim 7, characterized in that: When calculating the correlation between different features, the correlation layer uses the rectangular image patch imagelpatch of the source image to perform a convolution operation on the image2patch of the target image.

9. The SAR image adaptive registration method based on GLU-Net-Evaluate according to claim 8, characterized in that: When the final predicted optical flow field is generated by the optical flow estimation network and the image registration is completed by using the predicted optical flow field, the following steps are included: S31: first, extract features from the input target image It and the source image Is using a feature extractor at different resolutions, wherein the different resolutions include L1, L2, L3 and L4, and the resolution levels of L1, L2, L3 and L4 increase in sequence; S32: Then at resolution L1, the network calculates the global correlation between the source feature FL1 and the target feature FL1, and uses normalization and ReLU activation functions to process the obtained global correlation to strengthen the determined matches and reduce the weights of uncertain matches; S33: Global correlation is mapped to decoder M top is converted into a two-dimensional dense correspondence map m1, which is then converted into a displacement map (pixel coordinates of the image), and obtain the rough optical flow from the source image to the target image; S34: The calculated local correlation module is used on different resolution layers L2, L3 and L4 in turn to further refine the optical flow. At the same time, the optical flow decoder uses the local correlation and the optical flow sampled in the previous layer to calculate the residual optical flow, and then the residual optical flow is added to the optical flow of the previous layer to obtain the complete optical flow of the current layer.

10. The SAR image adaptive registration method based on GLU-Net-Evaluate according to claim 9, characterized in that: When the optical flow estimation network is used to generate the final predicted optical flow field and the predicted optical flow field is used to complete the image registration, the evaluation network is used to update and optimize the parameters of the optical flow estimation network. The evaluation network uses LSTM to generate a scene for evaluation and inputs it into the loss calculation function to guide the overall parameter update by integrating historical information. The workflow of the evaluation network is as follows: S41: receiving optical flow estimation data of the image as network input; S42: Network input data is processed through fully connected layers, batch normalization layers, and residual connections to achieve feature extraction and enhancement; S43: The extracted features are regularized by the Dropout layer to obtain a preliminary evaluation field; S44: LSTM layer processes the preliminary evaluation field; S45: The LSTM output passes through two fully connected layers to generate the evaluation field.

Images