A remote sensing image registration method based on a dual-channel attention convolutional network

Abstract

The application relates to a remote sensing image registration method based on a dual-channel attention convolutional network, which comprises the following steps: acquiring an image pair of a remote sensing data set to obtain a source image and a target image; adopting a dual-channel attention convolutional neural network to extract feature information of the input images respectively, and outputting extracted source image features S and target image features T, source image features S0 and target image features T0; respectively matching the feature maps of the dual channels, calculating the correlation of S and T, S0 and T0, and then using an improved feature matching method to match and screen two matching results to obtain a final result; obtaining image feature parameters through parameter regression and a grid loss function; and taking the image feature parameters as transformation parameters to guide the source image to perform affine transformation and complete image registration. The application solves the problems of low precision and low efficiency of multi-view and multi-temporal remote sensing image registration, and improves the precision and efficiency of multi-view and multi-temporal remote sensing topographic image registration.

Description

Technical Field

[0001] This invention relates to an image processing technique, and more particularly to a multi-view, multi-temporal remote sensing image registration method based on a dual-channel attention convolutional neural network. Background Technology

[0002] Multi-view, multi-temporal remote sensing terrain image registration is the process of matching and aligning pairs of remote sensing terrain images captured from the same or similar scenes under different viewpoints and time conditions using remote sensing camera tools, and finding the geometric correspondence between the image pairs. It is the foundation of visual tasks such as target recognition in remote sensing terrain images. How to improve the accuracy and efficiency of image registration is one of the hot topics in deep learning today.

[0003] Currently, the accuracy and feature extraction capabilities of neural networks remain bottlenecks limiting their development. In computer vision, shallow networks often extract geometric features such as image edges and textures, while deep networks extract higher-level features with semantic information. Existing deep feature extraction methods based on convolutional neural networks align the distribution of different data domains using a single feature structure. This approach results in only a portion of feature information being transferred from the source domain to the target domain. Summary of the Invention

[0004] To address the issues of low accuracy and efficiency in multi-view, multi-temporal remote sensing image registration, a remote sensing terrain image registration method based on a dual-channel attention convolutional neural network is proposed to improve the accuracy and efficiency of multi-view, multi-temporal remote sensing terrain image registration.

[0005] The technical solution of this invention is as follows:

[0006] A remote sensing image registration method based on a dual-channel attention convolutional network includes the following steps:

[0007] Step 1: Obtain the remote sensing image dataset, extract the source image and target image from the image pairs in the dataset, and use a generative adversarial network to transfer the image style of the target image to the source image to complete the data preprocessing;

[0008] Step 2: The two branches of the dual-channel attention convolutional neural network, ResNet101 and ResNet101 with added attention module, are used to extract the feature information of the input image, and the extracted source image features S and target image features T, source image features S0 and target image features T0 are output to the next feature matching layer.

[0009] Step 3: For the four feature maps obtained in Step 2, the feature maps of the two channels are matched respectively by the Euclidean distance matching method. The correlation between source image feature S and target image feature T, source image feature S0 and target image feature T0 is calculated. Then, the improved feature matching method is used to match and filter the two matching results to obtain the final result.

[0010] Step 4: Input the feature relationship obtained in Step 3 into the regression network to perform parameter regression, thereby obtaining a parameter, and input the parameter into the grid loss function to obtain the image feature parameters;

[0011] Step 5: Use the image feature parameters obtained in Step 4 as transformation parameters to guide the source image to perform affine transformation and complete image registration.

[0012] Furthermore, the migration operation in step one is as follows:

[0013] The target image B is input into a generative adversarial network to make it more similar to the source image A, resulting in the target image B1. Then, the source image A and the target image B1 are input into a convolutional neural network.

[0014] Furthermore, the dual-channel attention convolutional neural network in step two is specifically as follows:

[0015] The dual-channel attention-based neural network combines style transfer and adds a new ResNet101 network as a branch of the dual-channel network to the original pre-trained ResNet101 network, and improves it by combining attention mechanism.

[0016] The improved attention mechanism enhances the ResNet network by preserving low-level image features and extracting them through the attention mechanism before feeding them into the later stages of the network. This strengthens the fusion of low-level and high-level image features, retaining more information and ensuring that the extracted feature information is sufficiently rich, thereby further improving the accuracy of remote sensing image registration.

[0017] Furthermore, step two specifically involves:

[0018] First, the feature extraction network is pre-trained using a dataset. Then, the trained network is used to extract features from the input image. To better match the feature information output by the two channels and ensure sufficient fusion of high-level and low-level features, the fifth convolutional layer and all fully connected layers of the ResNet101 network are removed. A new ResNet101 channel is then added, incorporating an improved attention mechanism to form a dual-channel attention convolutional neural network. Each row of pixels in the remote sensing terrain image has different features; global context information for each row is obtained to generate the scaling factor X for each channel. L This can be used to estimate pixel-level channel weights during terrain image matching. To improve the fusion of low-level and high-level features, an improved attention mechanism is added after the third and fourth convolutional layers of the second network channel, ultimately resulting in an attention map along the channel direction. Where A represents the attention map. The network fuses all information representing the channel and vertical dimensions of the image, integrating image features; after calculating the attention map, the network extracts a high-level feature map X. h Through and New features obtained by element-wise multiplication It can be represented as:

[0019]

[0020] Meanwhile, in order to extract the contextual information of image features and the relationship between their channels, the attention mechanism uses wide pooling layers, interpolated coarse attention, and attention map computation steps; the structure of the dual-channel attention convolutional neural network consists of a ResNet101 network and an improved ResNet101 network, with the improved ResNet101 network adding an attention mechanism to the network.

[0021] Furthermore, the specific steps in step three regarding the attention mechanism, including the use of width-pooling layers, interpolated coarse attention, and attention graph computation, are as follows:

[0022] Width pooling is used to obtain channel attention maps, feature maps X. L G is obtained through width-wise pooling operation pool Extracting the contextual information of the height yields the matrix Z, which can be represented as: Z = G pool (X l The width pooling operation is empirically set to average pooling; formally, the h-th row vector of Z is calculated as:

[0023]

[0024] The coarse attention based on interpolation is the model-generated matrix after the merging operation. However, not all rows of matrix Z are useful for computing an effective attention map; downsampling the dimension C l ×H l The matrix Z interpolation is matrix This is a hyperparameter that is set empirically; since the attention map constructed from the downsampled feature map is coarse, it is transformed into one with a higher resolution than the given high-level feature map X through upsampling. h Same height dimension;

[0025] Attention map computation is the process of merging and interpolating feature maps in the width direction of the channel attention map A. The attention map A is obtained from the input convolutional layers; it indicates the important channels in each individual row; a sigmoid function is used when computing the attention map to allow for multiple labels; it can be composed of N convolutional layer operations, and can be represented as:

[0026]

[0027] Where σ is the sigmoid function and δ is the ReLU activation function. The i-th one-dimensional convolutional layer is defined.

[0028] Furthermore, step three specifically involves:

[0029] Based on the two network branches, the first-stage matching relationship between the source image S and the target image T, and between the source image S0 and the enhanced target image T0 are generated for the corresponding networks. Then, the second-stage optimal matching is performed to improve the registration accuracy. The two-stage feature matching is as follows:

[0030] In the first matching stage, at the feature matching layer, all feature coordinate positions f of the target image T are used respectively. T (i,j), each is compared with a certain feature f in the source image S. S (i k ,j k The process involves performing relevant mappings to form corresponding matching relationships. Then, ReLU activation, a fully connected layer, and sigmoid activation are used to channelize these matching relationships, thereby generating the corresponding correlation C from the source image S to the target image T. ST As shown in the following formula:

[0031] C ST (i,j,k)=f T (i,j) k f S (i k ,jk (4)

[0032] Similarly, in the feature matching layer, based on the coordinate positions of all features of the source image S... Each feature in the target image T is compared with a specific feature. Relevant mappings are performed to form corresponding matching relationships. Then, the matching relationships are channelized using the ReLU activation function, a fully connected layer, and the sigmoid activation function, thereby generating the corresponding matching relationship from the source image S to the target image T. As shown in the following formula:

[0033]

[0034] In the second matching stage, the inner product of the two extracted features is calculated, and then the optimal feature assignment matrix P is obtained according to the SinkhoRn algorithm. ij α i β j This ensures that P satisfies the three constraints of the allocation matrix, where λ is responsible for adjusting the degree of influence of information entropy, and M... ij Let S be the cost matrix. ij To obtain the optimal allocation matrix by substituting the correspondence into the SinkhoRn algorithm formula; the SinkhoRn algorithm formula and the feature matching vector inner product formula are as follows:

[0035]

[0036] .

[0037] Furthermore, step four specifically involves:

[0038] The image features obtained in step three are matched and input into the grid loss function to obtain the feature parameters L. Represents the image points at input. Let represent the output image points, l represent the grid loss function, and L represent the feature parameters obtained by summing the two pairs of matching relationships; the grid loss function is shown below:

[0039] .

[0040] Furthermore, step five specifically involves using the final feature parameters L obtained from the parameter regression in step four as the transformation parameter θ. resnet It is used to guide the source image to undergo affine transformation and complete image registration.

[0041] The beneficial effects of this invention are as follows:

[0042] Remote sensing image registration technology has broad application prospects in terrain detection, such as remote sensing image road change detection and remote sensing image terrain recognition.

[0043] This invention combines style transfer to make the image styles of the source image and the target image more similar. Then, a network branch ResNet101 with an improved attention mechanism is added to the original pre-trained ResNet101 network. The original ResNet101 network focuses on the local features of the image, while the other ResNet101 network with the improved attention mechanism focuses on the global features of the image and improves feature reuse.

[0044] This invention employs a two-stage matching method to improve feature matching. First, the features extracted from the two channels are matched separately, and then the matching results are used for a second-stage matching, which improves the registration accuracy. Attached Figure Description

[0045] Figure 1 This is a diagram of the algorithm framework of the present invention;

[0046] Figure 2 This is a diagram of the dual-channel attention convolutional network structure of the present invention;

[0047] Figure 3 This is a set of qualitative comparison experimental results of multi-view, multi-temporal mountain images according to the present invention;

[0048] Figure 4 This is a flowchart of the remote sensing image registration method of the present invention. Detailed Implementation

[0049] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0050] Step 1: In this embodiment, the Aerial Image Dataset (aviation image dataset) obtained from the paper AID: A Benchmark Dataset for Performance Evaluation of Aerial Scene Classification is used as the image input source. Source and target images are extracted from the image pairs in the dataset, and a generative adversarial network is used to transfer the image style of the target image to the source image, completing the data preprocessing.

[0051] The target image B is input into a generative adversarial network to make it more similar to the source image A in terms of color, texture, and other factors, resulting in the target image B1. Then, the source image A and the target image B1 are input into a convolutional neural network.

[0052] Due to the multi-temporal and multimodal nature of remote sensing images, significant differences in color, texture, and other factors often occur, leading to low registration accuracy. Therefore, generative adversarial networks (GANs) are used to preprocess remote sensing images to address this issue. This invention uses a GAN to perform adversarial generation on input image pairs, making their colors and textures more similar, thereby enhancing the accuracy of image registration.

[0053] Step 2: Feature extraction based on a dual-channel attention convolutional neural network

[0054] The described dual-channel attention-based neural network combines style transfer and adds a new ResNet101 network as a branch of the dual-channel network to the original pre-trained ResNet101 network, and improves it by combining attention mechanism.

[0055] First, a feature extraction network is pre-trained using the ImageNet dataset. Then, the trained network is used to extract features from the input image. To ensure better matching of the feature information output from the two channels and to fully integrate high-level and low-level features, this invention removes the fifth convolutional layer and all fully connected layers of the ResNet101 network. A new ResNet101 channel is then added, incorporating an improved attention mechanism to form a dual-channel attention convolutional neural network. Because remote sensing terrain images typically contain spatially based structural features, each row of image pixels has different features in terms of category distribution. Global context information for each row is obtained to generate the scaling factor X for each channel. L This can be used to estimate pixel-level channel weights during terrain image matching. To improve the fusion of low-level and high-level features, an improved attention mechanism is added after the third and fourth convolutional layers of the second network channel, ultimately resulting in an attention map along the channel direction. (A represents the attention map,) The network fuses image features (representing all information in the channel and vertical dimensions of the image). After computing the attention map, the network extracts a high-level feature map X. h Through and New features obtained by element-wise multiplication It can be represented as:

[0056]

[0057] To extract contextual information about image features and the relationships between their channels, the attention mechanism employs wide-pooling layers, interpolated coarse attention, and attention map computation. The dual-channel attention convolutional neural network consists of one ResNet101 network and one improved ResNet101 network. The improved ResNet101 network incorporates an attention mechanism, such as... Figure 2 As shown.

[0058] Width pooling is used to obtain the channel attention map, feature map X. L G is obtained through width-wise pooling operation pool Extracting the contextual information of the height yields the matrix Z, which can be represented as: Z = G pool (X l The width pooling operation is empirically set to average pooling. Formally, the h-th row vector of Z is calculated as:

[0059]

[0060] The coarse attention based on interpolation is the model-generated matrix after the merging operation. However, not all rows of matrix Z are useful for computing an effective attention map. Downsampling the dimension C... l ×H l The matrix Z interpolation is matrix This is a hyperparameter, empirically set to 16. Since the attention map constructed from the downsampled feature map is coarse, it is transformed by upsampling to have a higher resolution than the given high-level feature map X. h Same height dimension.

[0061] Attention map computation is the process of merging and interpolating feature maps in the width direction of the channel attention map A. The attention map A is obtained from the input convolutional layers. It indicates the important channels in each individual row. To allow for multiple labels, a sigmoid function is used instead of a softmax function when computing the attention map. It can be composed of N convolutional layer operations, and can be represented as:

[0062]

[0063] Where σ is the sigmoid function and δ is the ReLU activation function. The i-th one-dimensional convolutional layer is defined.

[0064] The improved attention mechanism of this invention improves the ResNet network by preserving the low-level features of the image and extracting them through the attention mechanism before feeding them into the later stages of the network. This enhances the fusion of low-level and high-level features of the image, preserving more information such as edges and shapes, and ensuring that the extracted feature information is rich enough, thereby further improving the accuracy of remote sensing image registration.

[0065] Step 3: Two-stage feature matching

[0066] This invention generates a first-stage matching relationship between the source image S and the target image T, and between the source image S0 and the enhanced target image T0, based on two network branches, and then performs a second-stage optimal matching to improve registration accuracy. The feature matching of the two stages is explained below.

[0067] In the first matching stage, at the feature matching layer, all feature coordinate positions f of the target image T are used respectively. T (i,j), each is compared with a certain feature f in the source image S. S (i k ,j k The process involves performing relevant mappings to form corresponding matching relationships. Then, ReLU activation, a fully connected layer, and sigmoid activation are used to channelize these matching relationships, thereby generating the corresponding correlation C from the source image S to the target image T. ST As shown below:

[0068] C ST (i,j,k)=f T (i,j) k f S (i k ,j k (4)

[0069] Similarly, in the feature matching layer, based on the coordinate positions of all features of the source image S... Each feature in the target image T is compared with a specific feature. Relevant mappings are performed to form corresponding matching relationships. Then, the matching relationships are channelized using the ReLU activation function, a fully connected layer, and the sigmoid activation function, thereby generating the corresponding matching relationship from the source image S to the target image T. As shown below:

[0070]

[0071] In the second matching stage, this invention calculates the inner product of the two extracted features, and then obtains the optimal feature allocation matrix P according to the SinkhoRn algorithm (optimal transport iteration algorithm). ij α i βj This ensures that P satisfies the three constraints of the allocation matrix, where λ is responsible for adjusting the degree of influence of information entropy, and M... ij Let S be the cost matrix. ij This is the optimal allocation matrix obtained by substituting the correspondence into the SinkhoRn algorithm formula. The SinkhoRn algorithm formula and the feature matching vector inner product formula are as follows:

[0072]

[0073]

[0074] Step Four:

[0075] The image features obtained in step three are matched and input into the grid loss function to obtain the feature parameters L. Represents the image points at input. Let represent the output image points, l represent the grid loss function, and L represent the feature parameters obtained by summing the two pairs of matching relationships. The grid loss function is shown below:

[0076]

[0077] Step 5:

[0078] The final feature parameters obtained from the parametric regression in step four will be used as the transformation parameter θ, with feature parameter L as the transformation parameter. resnet It is used to guide the source image to undergo affine transformation and complete image registration.

[0079] The deep learning algorithm proposed in this invention is trained and tested using publicly available aerial remote sensing image datasets as well as multi-view and multi-temporal remote sensing image datasets. It is compared with algorithms proposed in recent years, such as "A two-stream symmetric network with bidirectional ensemble for aerial image matching" and "Convolutional neural network architecture for geometric matching".

[0080] This paper is based on the PytoRch deep learning framework, using Python version 3.7, and the hardware environment consists of a GTX 3080 graphics card with 8GB of video memory and an Intel Core i7 (4.20GHz) processor.

[0081] Experimental Analysis and Comparison:

[0082] Figure 3 This paper presents the comparative experimental results of three algorithms—Two-stream Ensemble, CNN-GEO, and the algorithm proposed in this invention—on a set of multi-temporal and multi-view mountain images. When comparing registered and overlaid images, no difference in registration effect can be observed among the three methods. However, when comparing checkerboard details, the method proposed in this invention shows better registration effect in terms of checkerboard detail alignment, slightly better than the Two-stream Ensemble algorithm in terms of mountain road alignment, and significantly better than the CNN-GEO algorithm.

[0083] Table 1 shows the quantitative evaluation of the registration accuracy of the three algorithms using the Correct Keypoint Estimation (PCK) metric, the Root Mean Square Error (RMSE) metric, and the Time (TIME: in seconds) metric. The formulas for calculating PCK and RMSE are shown below:

[0084]

[0085] In the formula, α·max(h,w) represents a region of size α in an image region with height h and width w. For the key points in the source image obtained by transforming the i-th image pair, For the manually labeled key points in the target image of the i-th image pair, Here, d represents the distance between the transformed point and the original point within a certain range, and M represents the number of points.

[0086]

[0087] In the formula, M represents the pixel in the image, (x i ,y i (x′) represents the pixel obtained in the registered image. i ,y′ i () are manually marked pixels in the target image.

[0088] Table 1: Quantitative Comparison of Different Algorithms

[0089]

[0090] The higher the values ​​of PCK and RMSE, the higher the accuracy and the better the registration effect; the lower the value of time, the higher the efficiency of image registration. As shown in Table 1, the proposed method in this invention has the highest values ​​of PCK and RMSE and the lowest value of time, which effectively improves the accuracy and efficiency of multi-temporal remote sensing image registration to a certain extent.

[0091] This invention first incorporates style transfer by adding an improved ResNet101 network to the existing single ResNet101 network. This improves feature extraction from both global and local perspectives, enhances feature reuse, and more effectively fuses low-level and high-level feature information. Then, a two-stage feature matching process is used to obtain the correlation between the two networks. Finally, all parameters are obtained from the parameter regression network layers, and these parameters are synthesized with those from the original pre-trained network to obtain the final transformation parameters. These parameters guide the registration process of the source image through affine transformation. Experimental results show that the algorithm proposed in this invention effectively improves the accuracy and efficiency of remote sensing terrain image registration.

[0092] The above-described embodiments are merely one implementation of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention should be determined by the appended claims.

Claims

1. A remote sensing image registration method based on a dual-channel attention convolutional network, characterized in that, Includes the following steps: Step 1: Obtain the remote sensing image dataset, extract the source image and target image from the image pairs in the dataset, and use a generative adversarial network to transfer the image style of the target image to the source image to complete the data preprocessing; Step 2: The two branches of the dual-channel attention convolutional neural network, ResNet101 and ResNet101 with added attention module, are used to extract the feature information of the input image, and the extracted source image features S and target image features T, source image features S0 and target image features T0 are output to the next feature matching layer. Step 3: For the four feature maps obtained in Step 2, the feature maps of the two channels are matched respectively by the Euclidean distance matching method. The correlation between source image feature S and target image feature T, source image feature S0 and target image feature T0 is calculated. Then, the improved feature matching method is used to match and filter the two matching results to obtain the final result. Step 4: Input the feature relationship obtained in Step 3 into the regression network to perform parameter regression, thereby obtaining a parameter, and input the parameter into the grid loss function to obtain the image feature parameters; Step 5: Use the image feature parameters obtained in Step 4 as transformation parameters to guide the affine transformation of the source image and complete the image registration; The dual-channel attention convolutional neural network in step two is as follows: The dual-channel attention-based neural network combines style transfer and adds a new ResNet101 network as a branch of the dual-channel network to the original pre-trained ResNet101 network, and improves it by combining attention mechanism. The improved attention mechanism enhances the ResNet network by preserving low-level image features and extracting them through the attention mechanism before feeding them into the later stages of the network. This enhances the fusion of low-level and high-level image features, preserving more information and ensuring that the extracted feature information is rich enough, thereby further improving the accuracy of remote sensing image registration. Step two is as follows: First, the feature extraction network is pre-trained using a dataset. Then, the trained network is used to extract features from the input image. To better match the feature information output by the two channels and ensure sufficient fusion of high-level and low-level features, the fifth convolutional layer and all fully connected layers of the ResNet101 network are removed. A new ResNet101 channel is then added, incorporating an improved attention mechanism to form a dual-channel attention convolutional neural network. Each row of pixels in the remote sensing terrain image has different features; global context information for each row is obtained to generate the scaling factor X for each channel. L This can be used to estimate pixel-level channel weights during terrain image matching. To improve the fusion of low-level and high-level features, an improved attention mechanism is added after the third and fourth convolutional layers of the second network channel, ultimately resulting in an attention map along the channel direction. Where A represents the attention map, It represents all information in the channel and vertical dimensions of the image, and fuses the features of the image; After computing the attention map, the network extracts the high-level feature map X. h Through and New features obtained by element-wise multiplication , can be represented as: (1) Meanwhile, in order to extract the contextual information of image features and the relationship between their channels, the attention mechanism uses wide pooling layers, interpolated coarse attention, and attention map computation steps; the structure of the dual-channel attention convolutional neural network consists of a ResNet101 network and an improved ResNet101 network, with the improved ResNet101 network adding an attention mechanism to the network.

2. The remote sensing image registration method based on a dual-channel attention convolutional network according to claim 1, characterized in that, The migration operation in step one is as follows: The target image B is input into a generative adversarial network to make it more similar to the source image A, resulting in the target image B1. Then, the source image A and the target image B1 are input into a convolutional neural network.

3. The remote sensing image registration method based on a dual-channel attention convolutional network according to claim 1, characterized in that, Step three, which involves using a width-pooling layer, interpolated coarse attention, and attention map computation in the attention mechanism, specifically includes the following steps: Width pooling is used to obtain channel attention maps, feature maps X. L G is obtained through width-wise pooling operation pool Extracting the contextual information of the height yields the matrix Z, which can be represented as: Width pooling is empirically set to average pooling; formally, the h-th row vector of Z is calculated as: (2) The coarse attention based on interpolation is the model-generated matrix after the merging operation. However, not all rows of matrix Z are useful for computing an effective attention map; downsampling the dimension of The matrix Z interpolation is matrix ; This is a hyperparameter that is set empirically; since the attention map constructed from the downsampled feature map is coarse, it is transformed into one with a higher level of detail than the given high-level feature map through upsampling. Same height dimension; Attention map computation is the process of merging and interpolating feature maps in the width direction of the channel attention map A. The attention map A is obtained from the input convolutional layers; it indicates the important channels in each individual row; a sigmoid function is used when computing the attention map to allow for multiple labels; it can be composed of N convolutional layer operations, and can be represented as: (3) in It is the sigmoid function. It is the ReLU activation function. The i-th one-dimensional convolutional layer is defined.

4. The remote sensing image registration method based on a dual-channel attention convolutional network according to claim 1, characterized in that, Step three specifically involves: Based on the two network branches, the first-stage matching relationship between the source image S and the target image T, and between the source image S0 and the enhanced target image T0 are generated for the corresponding networks. Then, the second-stage optimal matching is performed to improve the registration accuracy. The two-stage feature matching is as follows: In the first matching stage, the feature matching layer matches the coordinates of all features of the target image T. Each feature in the source image S is compared with a specific feature. The relevant mapping is performed to form the corresponding matching relationship. Then, the ReLU activation function, fully connected layer and sigmoid activation function are used to channelize the above matching relationship, thereby generating the corresponding correlation relationship from the source image S to the target image T. As shown in the following formula: （4） Similarly, in the feature matching layer, based on the coordinate positions of all features of the source image S... Each feature in the target image T is compared with a specific feature. The process involves mapping the source image S to form a matching relationship, followed by channelization of the matching relationship using ReLU activation, a fully connected layer, and a sigmoid activation function. This process generates a corresponding matching relationship from the source image S to the target image T. As shown in the following formula: （5） In the second matching stage, the inner product of the two extracted features is calculated, and then the optimal feature assignment matrix is ​​obtained according to the SinkhoRn algorithm. ， Make Satisfying the three constraints of the allocation matrix, Responsible for adjusting the degree of influence of information entropy, For the cost matrix, To obtain the optimal allocation matrix by substituting the correspondence into the SinkhoRn algorithm formula; the SinkhoRn algorithm formula and the feature matching vector inner product formula are as follows: （6） （7）。 5. The remote sensing image registration method based on a dual-channel attention convolutional network according to claim 1, characterized in that, Step four is as follows: The image features obtained in step three are matched and input into the grid loss function to obtain the feature parameters L. Represents the image points at input. Let represent the output image points, l represent the grid loss function, and L represent the feature parameters obtained by summing the two pairs of matching relationships; the grid loss function is shown below: （8）。 6. The remote sensing image registration method based on a dual-channel attention convolutional network according to claim 5, characterized in that, Step five specifically involves using the final feature parameters L obtained from the parameter regression in step four as transformation parameters. It is used to guide the source image to undergo affine transformation and complete image registration.

Images