Agile satellite remote sensing image rapid registration method based on earth surface texture reference surface

Abstract

The invention discloses an agile satellite remote sensing image rapid registration method based on an earth surface texture datum plane, and particularly relates to the technical field of remote sensing image rapid registration. Geometric correction and radiation normalization are carried out on an image before registration, and then a representative texture sub-region is selected from each grid block of a reference image; extracting same structure features in the registered image in a sliding window mode, projecting the same structure features to a reference texture subspace for matching, analyzing error response for suppressing a repeated texture region by calculating a surface texture structure similarity coefficient, and further obtaining a large-angle side-sway imaging difference coefficient; the projection deformation degree of the registration image is quantitatively described from the geometric angle, a matching point drift superimposed effect model is constructed based on the projection deformation degree, the coupling risk relation between texture expression stability and imaging angle disturbance is revealed, and the matching point drift superimposed effect index is output based on the model. And intelligent scheduling and self-adaptive switching of the registration control strategy are realized.

Description

Technical Field

[0001] This invention relates to the field of rapid registration technology for remote sensing images, and more specifically, to a rapid registration method for agile satellite remote sensing images based on a surface texture reference surface. Background Technology

[0002] In rapid registration tasks using agile satellite remote sensing imagery, registration methods based on surface texture reference surfaces have gradually become a key approach for multi-temporal and multi-angle image registration due to their independence from ground feature geometry models and robustness to changes in illumination and radiation. However, in practical applications, the high repetition of surface texture structures and the imaging distortion caused by large-angle lateral tilting during agile imaging often exhibit a highly coupled influence mechanism, severely limiting the registration stability and accuracy of this method. In high-density urban scenes or regular farmland areas, although surface textures possess high energy, the strong repetition of texture forms such as building roofs, field segmentation, and road grids easily leads to non-unique local extrema in texture responses across multiple regions, resulting in fuzziness in texture matching. Simultaneously, agile satellites typically acquire target area images with large lateral tilting angles during trans-orbit imaging or high-revisit missions. When the lateral tilting angle changes significantly, severe non-affine projection distortion of the surface texture occurs, even leading to structural distortion or occlusion. In this scenario, repeating textures that already exhibit matching ambiguity will further lose their stable representation in the texture reference plane space due to changes in projection shape. This makes it difficult for sliding window matching or subspace mapping operations to establish a consistent mapping relationship, leading to the superposition of misaligned matching points and unstable registration parameters. Ultimately, this results in problems such as spatial jumps in registration errors, local stitching breaks, and global error accumulation. Therefore, this combined challenge of texture repetition and viewpoint distortion has become a key bottleneck restricting the widespread application of surface texture reference plane-based registration methods in complex urban scenes and large-area stitching tasks. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a rapid registration method for agile satellite remote sensing images based on a surface texture reference surface, in order to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: A fast registration method for agile satellite remote sensing images based on a surface texture reference surface includes the following steps: Step S1: Perform geometric correction and radiometric normalization on the Agile satellite remote sensing image, extract the effective surface area in the image, and divide the effective surface area into a structured grid according to a fixed resolution. Step S2: Select a texture sub-region within each grid block of the reference image to extract multi-scale directional texture features and construct a surface texture subspace for the texture sub-region. Step S3: In the registered image, a sliding window is used to traverse each region to be matched to extract texture features, and these features are projected into the surface texture subspace to obtain a preliminary matching score. Step S4: Obtain the surface texture structure similarity information of each grid block and its adjacent blocks in each sliding window matching area, and calculate the surface texture structure similarity coefficient. Step S5: Use imaging metadata to obtain the large-angle side-swing imaging difference information between the reference image and the registered image, and calculate the large-angle side-swing imaging difference coefficient. Step S6: Construct a texture matching point drift superposition effect model based on the surface texture structure similarity coefficient and the large-angle side-swing imaging difference coefficient, obtain the matching point drift superposition effect index, and assess the potential risk of extreme mismatch drift in the current texture matching. Step S7: Adaptively optimize the registration control strategy based on the potential risk of extreme mismatch drift in texture matching.

[0005] In a preferred embodiment, the agile satellite remote sensing imagery includes a reference image. Image registration .

[0006] In a preferred embodiment, a sub-region with high texture energy and structural stability is selected within each grid block of the reference image as the texture representative region of that grid: Gradient energy maps are used to detect texture-rich regions based on image regions within grid blocks. Non-maximum suppression is applied to the gradient energy map, and the K candidate regions with the largest responses are selected as texture sub-regions. ; For each texture sub-region A Gabor filter bank is used to extract multi-scale, multi-directional texture features to construct the surface texture representation: for each sub-region, feature vectors are extracted at S scales and D directions to form a texture feature matrix. : ,in The texture description dimension for a single region; Principal component analysis (PCA) was used to... Dimensionality reduction processing: Construct the surface texture subspace of the texture sub-region .

[0007] In a preferred embodiment, the regions in the registered image corresponding to the spatial positions of the reference image grid blocks are traversed by sliding a window, and multi-scale, multi-directional texture feature vectors are extracted for each sliding window. ; Texture feature vectors within the sliding window Perform a projection operation: ; After projection Projected samples of all known sub-regions in the i-th grid block of the reference image Compare and obtain an initial matching score: ,in For the initial matching score, The L2 norm of a vector; For each grid block Retain the top N candidate matching points with the smallest sliding window scores in the registered image.

[0008] In a preferred embodiment, by acquiring the surface texture structure similarity information of each grid block and its adjacent blocks, the strong texture artifacts existing in the surface texture structure are analyzed, and the surface texture structure similarity coefficient is calculated to measure the degree of strong texture artifacts existing in the surface texture structure. The logic for obtaining the surface texture structure similarity coefficient is as follows: Divide each grid block equally. sub-blocks For each sub-block Use a set of Gabor filters Perform convolution to extract its values ​​at different wavelength scales. The higher-order texture response tensor is obtained by analyzing the response characteristics at different orientation angles. : ,in For sub-blocks pixel values, Let the standard deviation be the Gaussian envelope. This is a convolution operation; Each sub-block Corresponding to a higher-order texture response tensor ; Each high-order texture response tensor After being flattened into a set of column vectors, they are mapped to a common texture subspace using multilinear SVD. : ,in For sub-blocks Vector representation in a unified common texture subspace; Each As graph nodes, construct an undirected graph. Node is The edge weights are the feature distances: ,in , Representing sub-blocks and sub-blocks Vector representation in a unified common texture subspace The standard deviation of the Gaussian envelope; The number of clusters after embedding the undirected graph according to the graph is set to . Graph clustering is performed on an undirected graph to obtain texture structure pattern partitioning, and intra-class aggregation degree is calculated. : ,in Let c be the set of sub-blocks contained in the c-th cluster. The number of sub-blocks, Euclidean distance; Calculate inter-class differences : ,in For clustering index, For the first Each cluster contains a set of sub-blocks; Calculate the similarity coefficient of surface texture structure : ,in This is a very small constant used to prevent division by zero.

[0009] In a preferred embodiment, by acquiring the large-angle side-swing imaging difference information between the reference image and the registered image, the drastic changes in the agile imaging viewing angle are analyzed, and the large-angle side-swing imaging difference coefficient is calculated to measure the drastic degree of the agile imaging viewing angle change. The logic for obtaining the difference coefficient of large-angle side-swing imaging is as follows: Pose parameters are extracted from the reference image and the registered image, and their respective imaging orientation vectors are constructed: ,in Here is the attitude rotation matrix. This is the standard vector pointing towards the ground in the satellite's body coordinate system. , These are the imaging direction vectors of the reference image and the registered image, respectively. Using imaging direction vector , Constructing the imaging cone for each image: Each pixel This directional vector penetrates to a fixed DEM elevation. Above, forming ground projection points; Select a set of overlapping grid blocks from the reference image and the registered image within the same geographic region. Based on the imaging direction vector, the position coordinates of the grid center pixel projected onto the ground DEM plane are obtained by projecting it onto the DEM surface. ; Calculate the residual between the two projections: ,in We obtain the tensor fields of all residuals: ,in This represents the number of overlapping grid blocks; Singular value decomposition of the residual tensor field: This yields the magnitudes of the three main directions of change; Calculate the difference coefficient of large-angle side-swing imaging : ,in This is an empirical scaling factor used to control the exponential convergence rate.

[0010] In a preferred embodiment, a texture matching point drift superposition effect model is constructed based on the surface texture structure similarity coefficient and the large-angle lateral tilt imaging difference coefficient to obtain the matching point drift superposition effect index. The formula used for the texture matching point drift superposition effect model is as follows: In the formula To match the superposition effect index of point drift, This represents the similarity coefficient of the surface texture structure. The difference coefficient for large-angle side-swing imaging. These represent the preset scaling factors for the surface texture structure similarity coefficient and the large-angle lateral tilt imaging difference coefficient, respectively. All are greater than 0.

[0011] In a preferred embodiment, the registration control strategy is adaptively optimized based on the potential risk of extreme mismatch drift in texture matching, as follows: If the superposition effect index of the matching point drift is greater than the threshold of the superposition effect index of the matching point drift, then the potential risk of extreme mismatch drift in the current texture matching is marked as a high potential extreme mismatch drift risk, and direct matching is suspended. If the superposition effect index of the matching point drift is less than or equal to the threshold of the superposition effect index of the matching point drift, then the potential risk of extreme mismatch drift in the current texture matching is marked as low potential extreme mismatch drift risk, and the standard feature matching process continues.

[0012] The technical effects and advantages of this invention are as follows: 1. This invention effectively shields the interference of grayscale inconsistencies caused by differences in sensor viewing angles and illumination variations by performing geometric correction and radiometric normalization on the image before registration, thus unifying the spatial scale and radiometric feature distribution of the registration input. Subsequently, representative texture sub-regions are selected from each grid block of the reference image, and stable directional texture features are extracted using a multi-scale directional filter. A low-dimensional texture subspace representation is constructed, making the feature distribution of the texture structure robust and generalizable. In the registered image, features with the same structure are extracted using a sliding window method and projected onto the reference texture subspace for matching, effectively avoiding structural mismatches caused by scale and angle changes in direct feature comparison. By calculating the similarity coefficient of the surface texture structure, erroneous responses in repetitive texture regions are analyzed and suppressed, improving the identification and discrimination stability of extreme matching points. Addressing the large-angle lateral distortion commonly found in agile satellite imaging, a large-angle lateral distortion imaging difference coefficient is further obtained to quantitatively characterize the projection deformation degree of the registered image from a geometric perspective. Based on this, a model of the drift and superposition effect of matching points is constructed, forming an evaluation mechanism for texture matching. The index criterion for extreme mismatch drift risk theoretically reveals the coupled risk relationship between texture representation stability and imaging angle perturbation. Based on the matching point drift superposition effect index output by this model, intelligent scheduling and adaptive switching of registration control strategies are realized, improving the robustness and accuracy of registration of agile remote sensing images in urban high-density areas, areas with repetitive texture distribution, and large-angle imaging conditions. It significantly reduces extreme mismatch rate and spatial drift error, and shows significant superiority in key indicators such as matching efficiency, control point quality, and large-scale stitching stability. It has extremely high engineering practical value and promotion prospects. Attached Figure Description

[0013] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings; Figure 1 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] Example: Figure 1 The present invention provides a fast registration method for agile satellite remote sensing images based on a surface texture reference surface, comprising the following steps: Step S1: Perform geometric correction and radiometric normalization on the Agile satellite remote sensing image, extract the effective surface area in the image, and divide the effective surface area into a structured grid according to a fixed resolution. Step S2: Select a texture sub-region within each grid block of the reference image to extract multi-scale directional texture features and construct a surface texture subspace for the texture sub-region. Step S3: In the registered image, a sliding window is used to traverse each region to be matched to extract texture features, and these features are projected into the surface texture subspace to obtain a preliminary matching score. Step S4: Obtain the surface texture structure similarity information of each grid block and its adjacent blocks in each sliding window matching area, and calculate the surface texture structure similarity coefficient. Step S5: Use imaging metadata to obtain the large-angle side-swing imaging difference information between the reference image and the registered image, and calculate the large-angle side-swing imaging difference coefficient. Step S6: Construct a texture matching point drift superposition effect model based on the surface texture structure similarity coefficient and the large-angle side-swing imaging difference coefficient, obtain the matching point drift superposition effect index, and assess the potential risk of extreme mismatch drift in the current texture matching. Step S7: Adaptively optimize the registration control strategy based on the potential risk of extreme mismatch drift in texture matching; In step S1, geometric correction and radiometric normalization are performed on the Agile satellite remote sensing image, as follows: The agile satellite remote sensing imagery includes reference imagery. Image registration ; Geometric correction: The image pixels (u,v) are mapped to ground coordinates (X,Y,Z) using the RPC model transformation function attached to the image, so as to unify them to the standard geographic coordinate system; Radiation normalization: A linear normalization method based on regional statistics for reference images Image registration Perform grayscale value alignment: ,in For the aligned and registered images, and These represent the standard deviations of pixel values ​​in the reference image and the registered image, respectively. and These represent the average pixel values ​​of the reference image and the registered image, respectively. The effective land surface region in the image is extracted using edge-preserving SLIC superpixel segmentation or semantic segmentation methods (such as the DeepLab series). Interference areas such as clouds, shadows, water bodies, and map edges are excluded. The effective land surface region is then divided into a structured grid at a fixed resolution, with each grid block denoted as […]. ; It should be noted that the edge-preserving SLIC superpixel segmentation method is an edge-preserving block clustering algorithm that can form texture-consistent regions while maintaining spatial continuity. Semantic segmentation methods (such as the DeepLab series) are based on deep convolutional neural networks and fuse dilated convolution with multi-scale contextual information to perform semantic segmentation of image regions. They are suitable for remote sensing semantic scene recognition. In this embodiment, the appropriate method can be flexibly selected according to image quality and processing requirements, and will not be elaborated here. Each grid block has the following structural properties: Geographic boundary coordinates: four corners ; Effective pixel mask ; In step S2, a sub-region with high texture energy and structural stability is selected within each grid block of the reference image as the texture representative region of that grid: Gradient energy maps are used to detect texture-rich regions based on image regions within grid patches. ,in This represents the texture energy response value. Let be the gradient of the image in the x-direction. This represents the gradient of the image in the y-direction. Non-maximum suppression is applied to the gradient energy map, and the K candidate regions with the largest responses are selected as texture sub-regions. ; For each texture sub-region Using Gabor filter banks: ,in Here is the formula for the two-dimensional function of the Gabor filter, where (x, y) are the coordinates of the original image. , The coordinates are the rotated coordinates, i.e., the original image coordinates along the direction angle. Rotated coordinates The wavelength is used to control the periodicity of the filter response, reflecting the sensitivity to a certain frequency. The unit is pixels; the smaller the value, the more sensitive it is to fine textures. The standard deviation of the Gaussian envelope controls the coverage of the filter in the spatial domain. A larger value results in a wider filtering range, while a smaller value leads to more localized filtering. The aspect ratio is used to control the "flatness" of the elliptical Gaussian envelope; Multi-scale and multi-directional texture features are extracted to construct the surface texture representation: For each sub-region, feature vectors are extracted at S scales and D directions to form a texture feature matrix. : ,in The texture description dimension for a single region; Principal component analysis (PCA) was used to... Dimensionality reduction processing: Construct the surface texture subspace of the texture sub-region The surface texture subspace is used to represent the main direction of surface texture change within the grid block; In step S3, the regions in the registered image corresponding to the spatial positions of the grid blocks in the reference image are traversed by sliding windows, and multi-scale, multi-directional texture feature vectors are extracted for each sliding window. ; It should be noted that multi-scale, multi-directional texture feature vectors can be extracted using Gabor filter response mean and energy, LBP (Local Binary Pattern) histogram, Sobel, Scharr and other directional gradient statistics. Texture feature vectors within the sliding window Perform a projection operation: ,in This represents the basis matrix obtained after training on each grid block of the reference image using PCA. The average texture feature of the i-th grid block in the reference image. The projected low-dimensional texture feature vector; After projection Projected samples of all known sub-regions in the i-th grid block of the reference image Compare and obtain an initial matching score: ,in For the initial matching score, The L2 norm of a vector; For each grid block Retain the top N candidate matching points with the smallest sliding window scores in the registered images; Step S4: Obtain the surface texture structure similarity information of each grid block and its adjacent blocks in each sliding window matching area, and calculate the surface texture structure similarity coefficient. In this invention, the surface texture structure similarity coefficient is a comprehensive metric used to measure the repeatability and directional consistency of the internal structure of surface texture regions in remote sensing images. Its core purpose is to identify and assess matching ambiguities caused by overly regular, repetitive, or highly directional surface textures, i.e., the potential interference of "strong texture artifacts" on registration accuracy. In actual remote sensing images, especially in high-density building areas, agricultural strips, and desert aeolian areas, although surface textures are rich in energy, they often have strong repeatability and high directional consistency. This causes multiple regions in the image to exhibit similar or even approximate response characteristics in texture space, forming the problem of "structural pseudo-uniqueness." In this case, if matching is based on texture similarity, multiple "extreme response points" may appear that are highly similar but mislocated, seriously interfering with the accurate positioning of matching points. The surface texture structure similarity coefficient introduced in this invention is precisely to solve this root cause problem of texture mismatch drift.

[0016] Specifically, a larger surface texture structure similarity coefficient indicates stronger similarity or regularity in the local texture structure of a certain area of ​​the image within the same grid block. This suggests the presence of significant texture artifacts in that area, which can easily lead to repetition of texture extrema, resulting in problems such as matching point drift, enhanced mismatches, and cumulative error amplification during matching. Conversely, a smaller surface texture structure similarity coefficient means that the texture in that area exhibits strong differences in structure, orientation, and texture distribution, indicating high discriminative power of texture features. In such areas, extrema are more unique, the spatial projection of matching points is relatively stable, and the registration process converges more easily. The surface texture structure similarity coefficient is not used in isolation but serves as one of the core inputs for constructing a texture matching point drift and superposition effect model. It works in conjunction with the large-angle imaging difference coefficient to model potential risks in the matching error space. The "matching point drift and superposition effect index," established through the combined effect of these two factors, can accurately identify the mismatch amplification effect caused by the combined effects of texture repetition and viewpoint distortion. Therefore, assessing the potential risk of extreme mismatch drift in current texture matching based on the surface texture structure similarity coefficient not only effectively characterizes the impact mechanism of texture ambiguity on matching accuracy at the theoretical level, but also provides a reliable index that is quantifiable, adjustable, and embeddable in practical operation. It has significant application value and engineering benefits for improving the automatic registration accuracy and robustness of agile satellite remote sensing images in complex scenarios such as large-view deformation and high-texture artifacts.

[0017] Therefore, by acquiring the surface texture structure similarity information of each grid block and its adjacent blocks, the strong texture artifacts existing in the surface texture structure are analyzed, and the surface texture structure similarity coefficient is calculated to measure the degree of strong texture artifacts existing in the surface texture structure. The logic for obtaining the surface texture structure similarity coefficient is as follows: Divide each grid block equally. sub-blocks For each sub-block Use a set of Gabor filters Perform convolution to extract its values ​​at different wavelength scales. The higher-order texture response tensor is obtained by analyzing the response characteristics at different orientation angles. : ,in For sub-blocks pixel values, Let the standard deviation be the Gaussian envelope. This is a convolution operation; Each sub-block Corresponding to a higher-order texture response tensor ; Each high-order texture response tensor After being flattened into a set of column vectors, they are mapped to a common texture subspace using multilinear SVD (HOSVD). : ,in For sub-blocks Vector representation in a unified common texture subspace; Each As graph nodes, construct an undirected graph. Node is The edge weights are the feature distances: ,in , Representing sub-blocks and sub-blocks Vector representation in a unified common texture subspace The standard deviation of the Gaussian envelope; The number of clusters after embedding the undirected graph according to the graph is set to . Graph clustering is performed on an undirected graph to obtain texture structure pattern partitioning, and intra-class aggregation degree is calculated. : ,in Let c be the set of sub-blocks contained in the c-th cluster. The number of sub-blocks, Euclidean distance; Calculate inter-class differences : ,in For clustering index, For the first Each cluster contains a set of sub-blocks; Calculate the similarity coefficient of surface texture structure : ,in This is to prevent division by zero by a very small constant (generally taken as...). ); It should be noted that the above formulas are all dimensionless calculations. Commonly used methods for removing dimensions include Min-Max normalization and Z-Score standardization, which will not be elaborated here. Step S5: Use imaging metadata to obtain the large-angle side-swing imaging difference information between the reference image and the registered image, and calculate the large-angle side-swing imaging difference coefficient. In this invention, the large-angle side-swing imaging difference coefficient is a metric used to quantify the differences in ground feature observations caused by changes in the agile imaging perspective between remote sensing images. Agile satellites, in order to achieve high-frequency, high-dynamic observation capabilities, often employ large-angle side-swing imaging techniques (such as ±30° or greater). While this improves spatiotemporal resolution efficiency, it also introduces significant differences in imaging angles. These differences can lead to nonlinear changes in the geometric structure, projection morphology, and even occlusion of the same surface texture in different images, especially noticeable in complex terrains such as urban high-rise buildings, mountain slopes, and the three-dimensional structure of bridges. This causes the registration point extrema to drift in texture space, directly inducing texture matching failure, error amplification, and local distortion. The larger the coefficient, the more drastic the difference in imaging perspective between the two images, the worse the consistency of the spatial morphology of the surface targets, and the less monotonic the projection mapping becomes. This makes it difficult to maintain the assumptions in the matching algorithm (such as affine and local continuity), significantly increasing the risk of drift or even mismatch at extreme response points. Conversely, a smaller coefficient indicates a more stable imaging perspective, more consistent local texture projection, higher robustness of extreme value localization, and easier establishment of a one-to-one correspondence between registration points in the surface texture space. By introducing a large-angle side-swing imaging difference coefficient and combining it with a surface texture structure similarity coefficient to construct a texture matching point drift superposition effect model, the potential mismatch drift risk can be predicted, effectively improving the robustness and adaptability of image registration in complex and agile imaging scenarios. Therefore, the introduction of this coefficient plays a crucial role in solving problems such as extreme value mismatch, matching drift, and multi-scale instability, especially in multi-temporal and multi-platform remote sensing big data registration scenarios, possessing significant engineering practical value and scientific innovation significance.

[0018] Therefore, by acquiring the large-angle side-swing imaging difference information between the reference image and the registered image, the drastic changes in the agile imaging viewpoint are analyzed, and the large-angle side-swing imaging difference coefficient is calculated to measure the drastic degree of the agile imaging viewpoint change. The logic for obtaining the difference coefficient of large-angle side-swing imaging is as follows: Pose parameters are extracted from the reference image and the registered image, and their respective imaging orientation vectors are constructed: ,in For attitude rotation matrix (heading) , up and down , roll ), This is the standard vector pointing towards the ground in the satellite's body coordinate system. , These are the imaging direction vectors of the reference image and the registered image, respectively. Using imaging direction vector , Constructing the imaging cone for each image: Each pixel This directional vector penetrates to a fixed DEM elevation. The top (DEM represents the elevation model of the Earth's surface (usually referring to the bare surface, excluding vegetation and buildings), is a two-dimensional geographic coordinate system, corresponding to the topographic elevation value of each location, forming ground projection points: ,in This is a directional projection function that outputs pixel points. Projected coordinates on the DEM plane This is the satellite's imaging position (orbit center point). From pixel The emitted imaging direction vector points towards the ground. The scaling factor of the ray passing through the DEM surface indicates the distance from the satellite position along... How much "distance" does it take to align with the DEM plane (height is...)? )intersect, This refers to the ground elevation of the DEM, i.e., the fixed DEM elevation set. ; Select a set of overlapping grid blocks from the reference image and the registered image within the same geographic region. Based on the imaging direction vector, the position coordinates of the grid center pixel projected onto the ground DEM plane are obtained by projecting it onto the DEM surface. ,in For reference image number The coordinates of the position of the center pixel of the grid projected onto the ground DEM plane , To register the image The coordinates of the position of the center pixel of the grid projected onto the ground DEM plane ; Calculate the residual between the two projections: ,in We obtain the tensor fields of all residuals: ,in This represents the number of overlapping grid blocks; Singular value decomposition (SVD) of the residual tensor field: This yields the magnitudes of the three main directions of change; Calculate the difference coefficient of large-angle side-swing imaging : ,in This is an empirical scaling factor used to control the exponential convergence rate; It should be noted that, The larger the value, the more sensitive it is to small geometric shifts, and the exponential term decreases rapidly; The smaller the value, the slower the function converges, and the less likely it is to trigger a high-difference response at small offsets. It can be set to be proportional to the inverse square of the ground projection resolution per unit pixel: It should be noted that the above formulas are all dimensionless calculations. Commonly used methods for removing dimensions include Min-Max normalization and Z-Score standardization, which will not be elaborated here. Step S6: Construct a texture matching point drift superposition effect model based on the surface texture structure similarity coefficient and the large-angle side-swing imaging difference coefficient, obtain the matching point drift superposition effect index, and assess the potential risk of extreme mismatch drift in the current texture matching. Based on the surface texture structure similarity coefficient and the large-angle lateral tilt imaging difference coefficient, a texture matching point drift superposition effect model is constructed to obtain the matching point drift superposition effect index. The formula used for the texture matching point drift superposition effect model is as follows: In the formula To match the superposition effect index of point drift, This represents the similarity coefficient of the surface texture structure. The difference coefficient for large-angle side-swing imaging. These represent the preset scaling factors for the surface texture structure similarity coefficient and the large-angle lateral tilt imaging difference coefficient, respectively. All are greater than 0; It should be noted that the above formulas are all dimensionless calculations. Commonly used methods for removing dimensions include Min-Max normalization and Z-Score standardization, which will not be elaborated here. The settings should be tailored to the specific circumstances. For example, an expert-empowered approach could be adopted, where experts in relevant fields are invited to determine the pre-defined proportions for each indicator through professional opinion surveys and comprehensive evaluations. It can be 0.5 or 0.5; As can be seen from the above calculation expressions, the larger the similarity coefficient of the surface texture structure and the larger the difference coefficient of the large-angle side-swing imaging, the larger the matching point drift superposition effect index. This indicates that the surface texture features of the current area have high repetition (i.e., strong texture artifacts exist). At the same time, the imaging angle has changed drastically, which significantly increases the risk of extreme mismatch and spatial drift of texture matching points during the registration process. Conversely, the smaller the similarity coefficient of the surface texture structure and the smaller the difference coefficient of the large-angle side-swing imaging, the smaller the matching point drift superposition effect index. This indicates that the surface texture features of the current area to be registered have good distinguishability and non-repetition. Moreover, the agile imaging angle change between the reference image and the registered image is relatively mild. The risk of extreme mismatch and spatial drift during the texture matching process is low. The registration algorithm runs more stably in this area, and the accuracy and reliability of the matching points are higher. Step S7: Adaptively optimize the registration control strategy based on the potential risk of extreme mismatch drift in texture matching, as follows: If the matching point drift superposition effect index is greater than the matching point drift superposition effect index threshold, it indicates that the current grid block has a high similarity in the surface texture structure (i.e., texture repetition or significant strong artifacts), and is accompanied by significant large-angle imaging perspective differences. Texture matching faces the dual interference risk of extreme value mismatch and position drift. The potential risk of extreme value mismatch drift in the current texture matching is marked as high potential extreme value mismatch drift risk. Direct matching is suspended, and pseudo-matching points are generated by using full-map geometric interpolation or regional interpolation estimation to prevent error propagation. If the matching point drift superposition effect index is less than or equal to the matching point drift superposition effect index threshold, it indicates that the current matching region has strong texture discriminability, and the imaging angle change is within an acceptable range. The matching stability is high, and the probability of mismatch and position drift is low. The potential risk of extreme mismatch drift in the current texture matching is marked as low potential extreme mismatch drift risk, and the standard feature matching process continues to be used. It should be noted that the threshold for the matching point drift and overlay effect can be set according to specific application requirements and the accuracy tolerance of the remote sensing image registration task. Common setting methods include: Statistical empirical method: Based on the distribution of the index value of the superposition effect of the drift of matching points in a large number of historical registration samples, the 95th percentile of the stable registration area is taken as the threshold. Dynamic adaptive method: Based on the mean and variance of the overall texture distribution and imaging angular difference distribution of the current image, the threshold level is adjusted in real time to improve adaptability.

[0019] This invention effectively shields against grayscale inconsistencies caused by sensor perspective differences and illumination variations by performing geometric correction and radiometric normalization on the image before registration, thus unifying the spatial scale and radiometric feature distribution of the registration input. Subsequently, representative texture sub-regions are selected from each grid block of the reference image, and stable directional texture features are extracted using a multi-scale directional filter. A low-dimensional texture subspace representation is constructed, making the feature distribution of the texture structure robust and generalizable. In the registered image, features with similar structures are extracted using a sliding window method and projected onto the reference texture subspace for matching, effectively avoiding structural mismatches caused by scale and angle variations in direct feature comparison. By calculating the surface texture structure similarity coefficient, erroneous responses in repetitive texture regions are analyzed and suppressed, improving the identification and discrimination stability of extreme matching points. Addressing the large-angle lateral distortion commonly found in agile satellite imaging, a large-angle lateral distortion imaging difference coefficient is further obtained to quantitatively characterize the projection deformation of the registered image from a geometric perspective. Based on this, a model of matching point drift and superposition effects is constructed, forming an evaluation mechanism for texture matching. The index criterion for extreme mismatch drift risk theoretically reveals the coupled risk relationship between texture representation stability and imaging angle perturbation. Based on the matching point drift superposition effect index output by this model, intelligent scheduling and adaptive switching of registration control strategies are realized, improving the robustness and accuracy of registration of agile remote sensing images in urban high-density areas, areas with repetitive texture distribution, and large-angle imaging conditions. It significantly reduces extreme mismatch rate and spatial drift error, and shows significant superiority in key indicators such as matching efficiency, control point quality, and large-scale stitching stability. It has extremely high engineering practical value and promotion prospects.

[0020] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0021] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0022] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A rapid registration method for agile satellite remote sensing images based on a surface texture reference surface, characterized by: Includes the following steps: Step S1: Perform geometric correction and radiometric normalization on the Agile satellite remote sensing image, extract the effective surface area in the image, and divide the effective surface area into a structured grid according to a fixed resolution. Step S2: Select a texture sub-region within each grid block of the reference image to extract multi-scale directional texture features and construct a surface texture subspace for the texture sub-region. Step S3: In the registered image, a sliding window is used to traverse each region to be matched to extract texture features, and these features are projected into the surface texture subspace to obtain a preliminary matching score. Step S4: Obtain the surface texture structure similarity information of each grid block and its adjacent blocks in each sliding window matching area, and calculate the surface texture structure similarity coefficient. Step S5: Use imaging metadata to obtain the large-angle side-swing imaging difference information between the reference image and the registered image, and calculate the large-angle side-swing imaging difference coefficient. Step S6: Construct a texture matching point drift superposition effect model based on the surface texture structure similarity coefficient and the large-angle side-swing imaging difference coefficient, obtain the matching point drift superposition effect index, and assess the potential risk of extreme mismatch drift in the current texture matching. Step S7: Adaptively optimize the registration control strategy based on the potential risk of extreme mismatch drift in texture matching.

2. The method for rapid registration of agile satellite remote sensing images based on a surface texture reference surface according to claim 1, characterized in that: The agile satellite remote sensing imagery includes reference imagery. Image registration .

3. The rapid registration method for agile satellite remote sensing images based on a surface texture reference surface according to claim 2, characterized in that: Within each grid block of the reference image, a sub-region with high texture energy and structural stability is selected as the texture representative region of that grid: Gradient energy maps are used to detect texture-rich regions based on image regions within grid blocks. Non-maximum suppression is applied to the gradient energy map, and the K candidate regions with the largest responses are selected as texture sub-regions. ; For each texture sub-region A Gabor filter bank is used to extract multi-scale, multi-directional texture features to construct the surface texture representation: for each sub-region, feature vectors are extracted at S scales and D directions to form a texture feature matrix. : ,in The texture description dimension for a single region; Principal component analysis (PCA) was used to... Dimensionality reduction processing: Construct the surface texture subspace of the texture sub-region .

4. The rapid registration method for agile satellite remote sensing images based on a surface texture reference surface according to claim 3, characterized in that: By sliding a window through the regions in the registered image corresponding to the spatial positions of the grid blocks in the reference image, multi-scale and multi-directional texture feature vectors are extracted for each sliding window. ; Texture feature vectors within the sliding window Perform a projection operation: ; After projection Projected samples of all known sub-regions in the i-th grid block of the reference image Compare and obtain an initial matching score: ,in For the initial matching score, The L2 norm of a vector; For each grid block Retain the top N candidate matching points with the smallest sliding window scores in the registered image.

5. The rapid registration method for agile satellite remote sensing images based on a surface texture reference surface according to claim 4, characterized in that: By acquiring the surface texture structure similarity information of each grid block and its adjacent blocks, the strong texture artifacts existing in the surface texture structure are analyzed, and the surface texture structure similarity coefficient is calculated to measure the degree of strong texture artifacts existing in the surface texture structure. The logic for obtaining the surface texture structure similarity coefficient is as follows: Divide each grid block equally. sub-blocks For each sub-block Use a set of Gabor filters Perform convolution to extract its values ​​at different wavelength scales. The higher-order texture response tensor is obtained by analyzing the response characteristics at different orientation angles. : ,in For sub-blocks pixel values, Let the standard deviation be the Gaussian envelope. This is a convolution operation; Each sub-block Corresponding to a higher-order texture response tensor ; Each high-order texture response tensor After being flattened into a set of column vectors, they are mapped to a common texture subspace using multilinear SVD. : ,in For sub-blocks Vector representation in a unified common texture subspace; Each As graph nodes, construct an undirected graph. Node is The edge weights are the feature distances: ,in , Representing sub-blocks and sub-blocks Vector representation in a unified common texture subspace The standard deviation of the Gaussian envelope; The number of clusters after embedding the undirected graph according to the graph is set to . Graph clustering is performed on an undirected graph to obtain texture structure pattern partitioning, and intra-class aggregation degree is calculated. : ,in Let c be the set of sub-blocks contained in the c-th cluster. The number of sub-blocks. Euclidean distance; Calculate inter-class differences : ,in For clustering index, For the first Each cluster contains a set of sub-blocks; Calculate the similarity coefficient of surface texture structure : ,in This is a very small constant used to prevent division by zero.

6. The rapid registration method for agile satellite remote sensing images based on a surface texture reference surface according to claim 5, characterized in that: By acquiring the large-angle side-swing imaging difference information between the reference image and the registered image, we analyze the drastic changes in the agile imaging viewpoint and calculate the large-angle side-swing imaging difference coefficient to measure the drastic degree of changes in the agile imaging viewpoint. The logic for obtaining the difference coefficient of large-angle side-swing imaging is as follows: Pose parameters are extracted from the reference image and the registered image, and their respective imaging orientation vectors are constructed: ,in Here is the attitude rotation matrix. This is the standard vector pointing towards the ground in the satellite's body coordinate system. , These are the imaging direction vectors of the reference image and the registered image, respectively. Using imaging direction vector , Constructing the imaging cone for each image: Each pixel This directional vector penetrates to a fixed DEM elevation. Above, forming ground projection points; Select a set of overlapping grid blocks from the reference image and the registered image within the same geographic region. Based on the imaging direction vector, the position coordinates of the grid center pixel projected onto the ground DEM plane are obtained by projecting it onto the DEM surface. ; Calculate the residual between the two projections: ,in We obtain the tensor field of all residuals: ,in This represents the number of overlapping grid blocks; Singular value decomposition of the residual tensor field: This yields the magnitudes of the three main directions of change; Calculate the difference coefficient of large-angle side-swing imaging : ,in This is an empirical scaling factor used to control the exponential convergence rate.

7. The rapid registration method for agile satellite remote sensing images based on a surface texture reference surface according to claim 6, characterized in that: Based on the surface texture structure similarity coefficient and the large-angle lateral tilt imaging difference coefficient, a texture matching point drift superposition effect model is constructed to obtain the matching point drift superposition effect index. The formula used for the texture matching point drift superposition effect model is as follows: In the formula To match the point drift superposition effect index, This represents the similarity coefficient of the surface texture structure. The difference coefficient for large-angle side-swing imaging. These represent the preset scaling factors for the surface texture structure similarity coefficient and the large-angle lateral tilt imaging difference coefficient, respectively. All are greater than 0.

8. The rapid registration method for agile satellite remote sensing images based on a surface texture reference surface according to claim 7, characterized in that: Based on the potential risk of extreme mismatch drift in texture matching, the registration control strategy is adaptively optimized as follows: If the superposition effect index of the matching point drift is greater than the threshold of the superposition effect index of the matching point drift, then the potential risk of extreme mismatch drift in the current texture matching is marked as a high potential extreme mismatch drift risk, and direct matching is suspended. If the superposition effect index of the matching point drift is less than or equal to the threshold of the superposition effect index of the matching point drift, then the potential risk of extreme mismatch drift in the current texture matching is marked as low potential extreme mismatch drift risk, and the standard feature matching process continues.

Images