Multi-modal remote sensing image deformation registration method and system
By fusing pre-trained models to extract structural coherence and global semantic features from multimodal remote sensing images, and utilizing channel attention mechanisms and neural displacement field decoders, the stability and computational cost issues of multimodal remote sensing image registration are resolved, achieving efficient and accurate image registration.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA TOWER CO LTD
- Filing Date
- 2026-01-13
- Publication Date
- 2026-06-12
Smart Images
Figure CN122199625A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of multimodal remote sensing image registration, and specifically relates to a method for deformable registration of multimodal remote sensing images. Background Technology
[0002] Multimodal remote sensing image registration is a crucial prerequisite for collaborative analysis of remote sensing data from different sensors in fields such as geographic monitoring and environmental research. Existing registration techniques mainly develop in two directions: one is the traditional method based on manually designed features, such as extracting local features like corners and edges of images through scale-invariant feature transformation algorithms for matching, or using cross-correlation calculations based on image patch grayscale information to find corresponding relationships; the other is the method based on deep learning, which uses architectures such as convolutional neural networks to automatically learn the feature representation of images and attempts to bridge the appearance differences between different modalities through models such as generative adversarial networks to improve the robustness of registration.
[0003] However, the aforementioned existing technologies still have significant limitations. Traditional handcrafted features lack stability when faced with imaging noise, lighting variations, and complex scenes, making them prone to mismatches and lacking an understanding of the global semantics of the image. Current deep learning methods often rely on feature extraction models trained on specific sensor data, resulting in a significant decrease in generalization ability when faced with unseen modalities. Furthermore, these methods typically require large amounts of precisely labeled data for training, leading to high computational costs, and the generated registration fields often fail to accurately capture subtle geometric deformations.
[0004] Therefore, how to achieve an efficient registration method that can adapt to multiple modalities, does not rely on dense annotations, and can accurately depict detailed deformations remains a technical problem that urgently needs to be solved in this field. Summary of the Invention
[0005] To address the aforementioned issues, this application provides a multimodal remote sensing image deformation registration method, which has the advantage of enhancing the model's adaptability to remote sensing images under different sensor types and imaging conditions.
[0006] This application provides a method for multimodal remote sensing image deformation registration, including acquiring source remote sensing images and target remote sensing images to be registered, and further including the following steps: Extract a first feature based on a first pre-trained model from the source remote sensing image, and extract a second feature based on a second pre-trained model from the target remote sensing image; The first feature and the second feature are combined to form a fused feature; The fused features are enhanced using a channel attention mechanism, and the enhanced features are output. The enhanced features are input into a neural network model that can generate displacement fields based on image features, and the dense displacement field between the source remote sensing image and the target remote sensing image is decoded and generated. Spatial transformation of the source remote sensing image is performed based on the dense displacement field to complete the registration with the target remote sensing image; The first pre-trained model is a pre-trained diffusion generation model used to extract features containing structural coherence, and the second pre-trained model is a pre-trained visual transformer model used to extract features containing global semantic information.
[0007] Furthermore, extracting the first feature includes: Extract multi-level features from the intermediate network layers of the diffusion generation model; Dimensionality reduction and upsampling are performed on multi-level features to obtain the first feature with uniform spatial resolution.
[0008] Furthermore, the diffusion generation model is a stable diffusion model, in which the intermediate network layer is an encoder-decoder structure, and the layers selected for extracting multi-level features include the second, fifth and eighth layers of this structure.
[0009] Furthermore, extracting the second feature includes: Extract the feature vectors corresponding to all image patches output from the last layer of the second pre-trained model, and use these as the second feature.
[0010] Furthermore, the fusion of the first feature and the second feature to form a fused feature includes: The first feature and the second feature are normalized separately. The normalized first feature and the second feature are concatenated along the channel dimension, and the contribution ratio of the two is controlled by the weight parameter to form a fused feature.
[0011] Furthermore, feature enhancement through channel attention mechanisms includes: The fusion features are initially refined to obtain the first intermediate feature; For the first intermediate feature, its channel-dimensional weights are adaptively recalibrated to obtain the channel-weighted feature; Spatial features are extracted from the channel-weighted features to obtain spatial features; A linear transformation along the channel dimension is applied to the spatial features to obtain the transformed features; The transformed features are added to the residuals of the first intermediate features to output the enhanced features.
[0012] Furthermore, the neural network model capable of generating displacement fields based on image features decodes and generates dense displacement fields by optimizing an objective function, which includes a feature alignment loss term based on the second norm and a smoothness regularization loss term based on the gradient of the dense displacement field.
[0013] This application also provides a multimodal remote sensing image deformation registration system, including an image acquisition module for acquiring the source remote sensing image and the target remote sensing image to be registered, and further including: The feature extraction module is used to extract a first feature based on a first pre-trained model and a second feature based on a second pre-trained model from the source remote sensing image and the target remote sensing image, respectively. The first pre-trained model is a pre-trained diffusion generation model used to extract features containing structural coherence, and the second pre-trained model is a pre-trained visual transformer model used to extract features containing global semantic information. The feature fusion module, connected to the feature extraction module, is used to fuse the first feature and the second feature to form a fused feature; The channel attention adaptation module is connected to the feature fusion module and is used to enhance the fused features and output the enhanced features. The neural displacement field decoding module, connected to the channel attention adaptation module, is used to decode and generate a dense displacement field between the source remote sensing image and the target remote sensing image based on the enhancement features. The image transformation and registration module is connected to the neural displacement field decoding module. It is used to perform spatial transformation on the source remote sensing image based on the dense displacement field to complete the registration with the target remote sensing image.
[0014] This application also provides an electronic device, which includes at least one processor and at least one memory, the memory being data-connected to the processor, wherein the memory stores instructions executable by at least one processor, the instructions being executed by at least one processor to enable at least one processor to perform any of the methods described above.
[0015] This application also provides a computer-storeable medium storing computer instructions, which, when executed by a processor, specifically perform the steps of any of the methods described above.
[0016] This application also provides a computer program product, including computer instructions, which, when executed by a processor, specifically perform the steps in any of the methods described above.
[0017] Compared with the prior art, this application has the following advantages: This invention employs a pre-trained diffusion generation model and a visual transformer model to extract features with structural coherence and global semantic information, and then fuses them. This results in features with rich semantics that are independent of specific imaging modalities, significantly enhancing the model's adaptability to remote sensing images under different sensor types and imaging conditions. It overcomes the shortcomings of poor robustness of traditional handmade features and weak generalization ability of existing deep learning methods.
[0018] Meanwhile, by introducing a channel attention adapter to enhance the fused features, the feature channels that are crucial to the registration task can be adaptively highlighted, while irrelevant or interfering information is suppressed, thereby improving the quality and discriminative power of feature representation. Furthermore, a neural displacement field decoder is designed to directly infer the dense displacement field based on the enhanced features, avoiding the complex iterative optimization process in traditional methods. This significantly reduces computational costs while ensuring registration accuracy, achieving efficient processing.
[0019] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description
[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0021] Figure 1 A flowchart of a method according to an embodiment of this application is shown; Figure 2 A system structure diagram according to an embodiment of this application is shown. Detailed Implementation
[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0023] This invention is mainly applied to a wide range of fields that require the fusion and analysis of multi-source Earth observation data. Typical application scenarios include, but are not limited to, map updating and production in geographic information systems, land use change monitoring, forest cover assessment and water pollution tracking in environmental science research, building change detection and infrastructure layout analysis in urban planning, flood range mapping and damage assessment in disaster emergency response, and crop growth monitoring and yield forecasting in the agricultural field.
[0024] In the above scenarios, the data usually comes from satellites or airborne platforms equipped with different physical sensors, such as visible light / multispectral cameras, synthetic aperture radar, hyperspectral imagers, and infrared sensors. This results in remote sensing image pairs with different modalities, complementary information, but also significant differences. High-precision spatial registration of these multimodal images is the cornerstone for effectively associating, comparing, and fusing information from different data sources. It is an indispensable preprocessing step to improve the reliability of subsequent analysis and decision-making.
[0025] To address the aforementioned issues, this invention integrates features from two pre-trained base models with complementary characteristics to obtain semantically rich and modality-independent general feature representations, thereby improving the model's adaptability and generalization ability to images from different sensors. Furthermore, it designs a channel attention adapter and a neural network model (such as a neural displacement field decoder) capable of generating displacement fields based on image features. The former adaptively enhances key feature channels, suppresses noise, and improves feature quality, while the latter efficiently generates accurate displacement fields with a single forward inference, avoiding complex iterative optimization. The entire scheme employs self-supervised learning, requiring no additional labeled data, significantly reducing the application threshold and cost.
[0026] Please see Figure 1 The specific methods for achieving the above effects will be explained in detail below.
[0027] S1. Feature extraction and fusion based on pre-trained models.
[0028] The input for this step is a set of multimodal remote sensing images to be registered, including a target remote sensing image as a reference. and a source remote sensing image that needs to be aligned with it. These two images can come from different sensors, such as For optical images, and This is a synthetic aperture radar image.
[0029] S1-1, Extract features of the diffusion generation model (first feature).
[0030] Source remote sensing images The input is fed into a pre-trained diffusion generative model (first pre-trained model) to extract features containing structural coherence. In this embodiment, the diffusion generative model is a stable diffusion model, which is a latent diffusion model whose internal U-Net structure (encoder-decoder structure) intermediate layers learn rich prior knowledge of visual concepts and spatial structures.
[0031] Specifically, feature maps are extracted from and combined from the second, fifth, and eighth layers of the encoder-decoder path of the diffusion generation model. This choice is made because shallower layers (such as the second layer) tend to capture more global semantic information, while deeper layers (such as the eighth layer) preserve finer low-level textures and appearance details, and the middle layer (the fifth layer) serves as a transition.
[0032] Subsequently, principal component analysis was performed on the original features extracted from these three layers (in some embodiments, the total number of channels is 2720) to reduce dimensionality and remove redundancy. Then, the dimensionality-reduced features of each layer were upsampled to the same spatial resolution and finally concatenated to obtain the first feature representing structural coherence, denoted as . .
[0033] S1-2, Extracting visual transformer model features (second feature): remote sensing images of the target The input is fed into a pre-trained visual transformer model (second pre-trained model) to extract features containing global semantic information. In this embodiment, the visual transformer model is a DINOv2 model pre-trained based on self-supervised contrastive learning. This model can model the long-range dependencies between all pixel blocks in the image through a self-attention mechanism, thereby capturing global contextual semantic information.
[0034] Specifically, feature representations corresponding to all image patches output from the last layer of the visual transformer model are extracted. These representations are used as a highly abstract second feature that encodes the global semantic content of the image, denoted as... In this embodiment, The dimension is 768.
[0035] S1-3, Feature fusion; S1-3-1, Obtain and Afterwards, and Layer normalization or batch normalization is performed separately to stabilize training and accelerate convergence.
[0036] because It is a spatial feature map (e.g., with dimensions H×W×C1), while Typically, it's a global vector (e.g., 1×1×C2) or something that can be transformed into a global vector, requiring... The device is copied and expanded in spatial dimensions to make its size H×W×C2, so as to be compatible with... Perform point-by-point fusion.
[0037] S1-3-2. The two spatially aligned feature sets are concatenated along the channel dimension, and a learnable weight parameter is introduced. This is to balance the relative importance of the two types of features. The formula for calculating the fused feature Fc is as follows:
[0038] in, 2 indicates a splicing operation along the channel dimension.
[0039] in, This means that during model training, automatic learning is achieved through backpropagation, enabling the network to adaptively adjust its dependence on structural detail features and global semantic features based on different input image pairs. The resulting feature Fc possesses both detail resolution and semantic consistency, providing a powerful feature foundation for dealing with complex multimodal differences.
[0040] S2, Feature enhancement is performed using channel attention mechanism.
[0041] Since directly fused features may contain noise or redundant information, this step processes the fused features through a channel attention mechanism to further improve feature quality and highlight channels that are crucial to the registration task.
[0042] The channel attention mechanism can be implemented through a channel attention adapter, the working process of which includes feature extraction, adaptive calibration of channel weights, spatial feature extraction, and residual fusion.
[0043] In a specific embodiment, the channel attention adapter consists of sequentially connected aggregation blocks and feature enhancement blocks.
[0044] S2-1, the aggregation block performs preliminary refinement of the input fused features. In a specific implementation, it is built based on standard residual blocks and contains multiple convolutional layers and a shortcut connection. This structure can effectively extract contextual information while preserving the original features. Its output is denoted as... .
[0045] S2-2, Feature enhancement block for refined features The core of this enhancement is the introduction of a channel attention mechanism to adaptively recalibrate the weights of the feature channels. In a specific implementation, the steps are as follows: S2-2-1. Global channel description is obtained by global average pooling through the squeezing excitation module. S2-2-2, Subsequently, the global channel description is stimulated through two fully connected layers to generate adaptive weights for each channel, thereby highlighting important channels and suppressing secondary channels to obtain channel-weighted features; S2-2-3. Spatial feature extraction is performed on the weighted features after passing through a deep convolutional layer.
[0046] S2-2-4. A pointwise convolutional layer is used to perform a linear transformation on the channel dimension of the extracted spatial features. The transformation result is then added to the input features of the feature enhancement block through a residual connection to output the enhanced feature Fen, as shown in the following formula:
[0047] in, Indicates the excitation block being squeezed. Representation layer normalization, Represents depthwise convolution. This represents pointwise convolution.
[0048] S3. Generate the displacement field based on the neural network model (neural displacement field decoder).
[0049] The goal of this step is to use the optimized enhanced features of the source and target remote sensing images obtained in step S2 as input to regress the features from... and dense displacement field .
[0050] After training, the neural displacement field decoder can directly infer the displacement field from the aforementioned enhanced features, thereby avoiding the complex iterative optimization process in traditional registration methods and improving computational efficiency.
[0051] This step includes: S3-1. Construct a neural network model that can generate a displacement field based on image features, namely a neural displacement field decoder. This model takes the enhanced features as input and outputs a dense displacement field from the source image to the target image.
[0052] S3-2. To ensure the accuracy of the displacement field, a feature alignment loss function is defined. Its core principle is that a correct displacement field should make the transformed source image as similar as possible to the target image in the feature space. This loss is calculated by evaluating the enhanced features of the source image. According to the current predicted displacement field Features obtained after sampling the target image The difference between them is measured, specifically using a measure based on the second norm (L2 norm), the formula of which is:
[0053] in, The feature enhancement function used by the channel attention adapter represents the feature enhancement function for the channel attention adapter. Extract its enhanced features; The feature enhancement function used by the channel attention adapter represents the feature enhancement function for the channel attention adapter. Extract its enhanced features.
[0054] Minimizing this loss drives the model to predict the displacement field that enables the two images to be precisely aligned in terms of high-level semantic and structural features.
[0055] S3-3. To ensure the physical plausibility of the displacement field, a smoothness regularization loss function is defined. Relying solely on feature alignment loss, the model may predict physically abrupt or discontinuous displacement fields, leading to distortion in the registered image. The smoothness regularization loss function constrains the spatial gradient of the predicted displacement field, encouraging gentle and gradual changes in the displacement field, consistent with the continuous deformation characteristics of most natural scenes. The smoothness regularization loss function is:
[0056] S3-4. Construct the total loss function for model training. The total loss function is a weighted sum of the feature alignment loss and the smoothness regularization loss, used to comprehensively guide model learning. Its formula is:
[0057] in, is a hyperparameter used to balance feature alignment and gradient-based smoothness regularization; This represents the total loss.
[0058] S3-5, Neural Network Model Training.
[0059] Since the training of the neural network model (neural displacement field decoder) is self-supervised and does not require any manually labeled geometric correspondence information, the neural network model can be trained by minimizing the total loss function using the standard backpropagation algorithm.
[0060] S3-6, Application of Neural Network Models and Displacement Field Generation.
[0061] After training, the model has the ability to directly infer the displacement field from the aforementioned enhanced features, thus avoiding complex iterative optimization in the application stage.
[0062] During deployment, the enhanced features obtained in step S2 are input into the trained neural network model to decode and generate a dense displacement field between the source remote sensing image and the target remote sensing image. .
[0063] S4, Image Transformation and Registration Output.
[0064] S4-1. Use the dense displacement field obtained from decoding to perform spatial transformation on the source remote sensing image.
[0065] S4-2. Using differentiable bilinear sampling and other methods, the image is resampled according to the displacement vector indicated by each pixel to obtain the registered image.
[0066] S4-3. Complete the precise spatial alignment of the source remote sensing image and the target remote sensing image, and output the aligned image.
[0067] Please see Figure 2 Corresponding to the foregoing method claims, the present invention also provides a multimodal remote sensing image deformation registration system. This system implements the complete process of the above method in a modular manner, with each module connected sequentially to form a data processing pipeline. The core components of the system include: The image acquisition module, as the data entry point, is responsible for inputting the source remote sensing image and the target remote sensing image to be registered; The feature extraction module, connected to the image acquisition module, has the core function of performing the key pre-trained model feature extraction step in the aforementioned method. Specifically, this module is configured to call and run a pre-trained diffusion generation model (such as a stable diffusion model) to extract the first feature containing structural coherence from the input image, and at the same time call and run a pre-trained visual transformer model (such as the DINOv2 model) to extract the second feature containing global semantic information. The feature fusion module, whose input is connected to the output of the feature extraction module, is responsible for receiving the first feature and the second feature mentioned above, and performing feature normalization, channel dimension concatenation and learnable weight parameter weighting operations to generate a unified fused feature. The function of this module corresponds to the detailed calculation process of feature fusion in the method. The channel attention adaptation module connects to the output of the feature fusion module at its input end. This module has built-in the aggregation block and feature enhancement block as described above. By aggregating the input fused features and adaptively adjusting the channel weights based on the squeezing excitation mechanism, it outputs enhanced features with higher quality and greater discriminative power. The neural displacement field decoding module receives enhanced features from the channel attention adaptation module as input. This module is essentially a lightweight convolutional neural network. It directly infers the dense displacement field from the source remote sensing image to the target remote sensing image by optimizing the objective function that combines feature alignment loss and smoothness regularization loss, thus completely avoiding iterative optimization. The image transformation and registration module connects to a neural displacement field decoding module at its input to obtain the dense displacement field. This module integrates differentiable spatial transformation functions (such as a bilinear sampler) and uses this displacement field to geometrically deform the source remote sensing image, ultimately outputting a result image that is accurately registered with the target remote sensing image. All of the above modules can be implemented in a computing device through software, hardware, or a combination of both. They work together to automate the process from multimodal image input to high-precision registration result output.
[0068] Based on the same inventive concept, the present invention also provides an electronic device. This electronic device may be a server, personal computer, workstation, or dedicated image processing device, etc., and includes at least one processor (e.g., a central processing unit (CPU), graphics processing unit (GPU), or tensor processing unit (TPU)) and at least one memory (e.g., random access memory (RAM), read-only memory (ROM), or flash memory). The memory and the processor are connected via a system bus or other means, wherein the memory stores computer program instructions executable by the at least one processor. When the electronic device is running, the processor reads and executes the instructions stored in the memory, thereby controlling the electronic device to perform various steps in any of the aforementioned multimodal remote sensing image deformation registration methods, such as controlling the data flow to complete the entire process from image pair input, feature extraction and fusion based on a pre-trained model, channel attention feature enhancement, neural displacement field decoding to the final image transformation output.
[0069] Furthermore, the present invention provides a computer-readable storage medium. This storage medium can be a non-transitory tangible medium, such as a disk, optical disk, solid-state drive, USB flash drive, or memory card, on which computer program instructions are stored. When the computer instructions in this storage medium are loaded into a processor such as the aforementioned electronic device and executed, they specifically guide the processor to perform all operational steps in any of the multimodal remote sensing image deformation registration methods described above. These instructions constitute the software carrier for implementing the technical solution of the present invention.
[0070] Similarly, the present invention also provides a computer program product. This program product can be directly embodied as a set of computer instructions stored on a storage medium, or as a data packet downloaded via a network. The computer instructions contained in the computer program product, when executed by the processor of a device, will specifically perform the steps in any of the aforementioned multimodal remote sensing image deformation registration methods. This program product is another manifestation of the technical solution of the present invention, and its purpose is to enable the method of the present invention to be implemented and applied on different computing platforms through software distribution and operation.
[0071] The above descriptions of the system, equipment, media, and program products, along with the descriptions of the foregoing method embodiments, are based on the same technical principles and inventive concepts. Together, they constitute a complete implementation system of the technical solution of this invention. Those skilled in the art will understand that various modifications and variations can be made to the system, equipment, their module connections, and software implementation details described herein without departing from the spirit and scope of this invention.
[0072] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A method for deformable registration of multimodal remote sensing images, comprising acquiring a source remote sensing image and a target remote sensing image to be registered, characterized in that, It also includes the following steps: Extract a first feature based on a first pre-trained model from the source remote sensing image, and extract a second feature based on a second pre-trained model from the target remote sensing image; The first feature and the second feature are combined to form a fused feature; The fused features are enhanced using a channel attention mechanism to output enhanced features; The enhanced features are input into a neural network model that can generate a displacement field based on image features, and a dense displacement field between the source remote sensing image and the target remote sensing image is decoded and generated. The source remote sensing image is spatially transformed based on the dense displacement field to complete the registration with the target remote sensing image; The first pre-trained model is a pre-trained diffusion generation model used to extract features containing structural coherence, and the second pre-trained model is a pre-trained visual transformer model used to extract features containing global semantic information.
2. The method according to claim 1, characterized in that, Extracting the first feature includes: Extract multi-level features from the intermediate network layers of the diffusion generation model; The multi-level features are subjected to dimensionality reduction and upsampling to obtain the first feature with uniform spatial resolution.
3. The method according to claim 2, characterized in that, The diffusion generation model is a stable diffusion model, in which the intermediate network layer is an encoder-decoder structure, and the layers selected for extracting multi-level features include the second, fifth and eighth layers of this structure.
4. The method according to claim 1, characterized in that, Extracting the second feature includes: Extract the feature vectors corresponding to all image patches output from the last layer of the second pre-trained model, and use these as the second feature.
5. The method according to claim 1, characterized in that, The fusion of the first feature and the second feature to form a fused feature includes: The first feature and the second feature are normalized respectively; The normalized first feature and the second feature are concatenated along the channel dimension, and the contribution ratio of the two is controlled by the weight parameter to form the fused feature.
6. The method according to claim 1, characterized in that, The feature enhancement via channel attention mechanism includes: The fusion features are initially refined to obtain the first intermediate feature; For the first intermediate feature, its weights in the channel dimension are adaptively recalibrated to obtain the channel-weighted feature; Spatial features are extracted from the channel-weighted features to obtain spatial features; A linear transformation along the channel dimension is applied to the spatial features to obtain the transformed features; The transformed feature is added to the residual of the first intermediate feature to output the enhanced feature.
7. The method according to claim 1, characterized in that, A neural network model capable of generating displacement fields based on image features decodes and generates the dense displacement field by optimizing an objective function, which includes a feature alignment loss term based on the second norm and a smoothness regularization loss term based on the gradient of the dense displacement field.
8. A multimodal remote sensing image deformation registration system, comprising an image acquisition module for acquiring a source remote sensing image and a target remote sensing image to be registered, characterized in that, Also includes: The feature extraction module is used to extract a first feature based on a first pre-trained model and a second feature based on a second pre-trained model from the source remote sensing image and the target remote sensing image, respectively. The first pre-trained model is a pre-trained diffusion generation model used to extract features containing structural coherence, and the second pre-trained model is a pre-trained visual transformer model used to extract features containing global semantic information. A feature fusion module, connected to the feature extraction module, is used to fuse the first feature and the second feature to form a fused feature; The channel attention adaptation module is connected to the feature fusion module and is used to enhance the fused features and output the enhanced features. A neural displacement field decoding module, connected to the channel attention adaptation module, is used to decode and generate a dense displacement field between the source remote sensing image and the target remote sensing image based on the enhanced features. An image transformation and registration module, connected to the neural displacement field decoding module, is used to perform spatial transformation on the source remote sensing image based on the dense displacement field, thereby completing the registration with the target remote sensing image.
9. An electronic device, characterized in that, The electronic device includes at least one processor and at least one memory, the memory being data-connected to the processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-7.
10. A computer-storable medium, characterized in that, The storable medium stores computer instructions, which, when executed by a processor, specifically perform the steps of the method as described in any one of claims 1-7.
11. A computer program product comprising computer instructions, characterized in that, When the computer instructions are executed by the processor, they specifically perform the steps in the method as described in any one of claims 1-7.