A large-scale SAR image registration method and system based on irregular sub-regions

Abstract

This invention relates to the field of synthetic aperture radar (SAR) image registration technology, specifically to a method and system for large-scale SAR image registration based on irregular regional segmentation. The method includes: segmenting two SAR images of the same area at different times into superpixel blocks based on preset multimodal rules; dynamically merging adjacent blocks according to the regional similarity of the superpixel blocks to obtain superpixel regions; determining the polynomial order of the adaptive local migration model based on the deformation complexity and feature point distribution of the superpixel regions; performing global fusion optimization and region merging optimization based on model parameters to obtain optimized superpixel regions; and cropping the common areas of the optimized regions to obtain the registration result. The system includes: superpixel segmentation, dynamic merging, model order determination, region optimization, and cropping units. This invention takes into account both the adaptability to local distortion of complex large-area SAR images and the consistency of global registration, and is particularly suitable for scenes with severe geometric deformation such as mountains and coastlines.

Description

Technical Field

[0001] This invention belongs to the field of synthetic aperture radar image registration technology, specifically relating to a method and system for large-scale SAR image registration based on irregular regional division. Background Technology

[0002] SAR (Synthetic Aperture Radar) is an active Earth observation system that can be mounted on aircraft, satellites, spacecraft, and other flying platforms. It can conduct all-weather, 24 / 7 Earth observation and has a certain degree of surface penetration capability. SAR images are susceptible to geometric distortion due to side-looking imaging characteristics and speckle noise, leading to significant geometric and radiometric differences in the same target across different images. SAR image registration is a technique that aligns two or more images acquired at different times, from different viewing angles, or from different sensors to the same coordinate system. Its core function is to eliminate spatial bias, providing a spatial consistency basis for tasks such as SAR image fusion, interferometry, temporal analysis, and target detection. SAR image registration methods can be categorized by their technical principles into intensity-based, feature-based, deep learning-based, and hybrid registration methods, each with its unique advantages and limitations.

[0003] Intensity-based registration methods are common techniques that directly utilize pixel intensity or statistical characteristics to achieve image alignment. For example, the mutual information maximization method optimizes spatial transformation parameters by statistically analyzing the gray-level distribution dependence of two images. It is suitable for multi-temporal or multi-polarization data registration, but it is sensitive to speckle noise and easily gets trapped in local optima. Coherence coefficient optimization methods, based on interferometric SAR data, achieve high-precision registration through interferometric coherence measurements, but they depend on the quality of the interferometric data and cannot be applied to non-interferometric SAR data scenarios. These methods do not require feature extraction, but the computation time increases significantly with image size, making it difficult to meet the needs of efficient processing over large areas.

[0004] Feature-based registration methods achieve matching by extracting stable structural features from images. The SAR-SIFT algorithm, based on the traditional SIFT, is optimized for the radiometric characteristics and speckle noise of SAR images. It generates matching point pairs using strong scatterers or edge features, and is suitable for areas with prominent ground features. However, it is prone to failure in weakly textured environments due to insufficient features. The improved Harris-SSC method combines Harris corner detection and structural similarity constraints, and improves stability by selecting high-confidence matching points. However, it has high computational complexity, requires large-area images to be processed in blocks, and is prone to causing inconsistencies in deformation between blocks.

[0005] In recent years, deep learning-based registration methods have developed rapidly. SAR-CNN learns the nonlinear deformation field between images through convolutional networks and outputs registration parameters end-to-end, enabling it to adapt to complex geometric distortions. However, it relies on a large amount of labeled data and has weak cross-scene application capabilities. Transformer-based models utilize self-attention mechanisms to capture long-range spatial dependencies, theoretically making them more suitable for large-area global consistency modeling. However, due to limitations in GPU memory and computing resources, they face real-time challenges in practical applications. These methods excel in noise suppression and deformation modeling, but they have high hardware requirements, and the issue of GPU memory consumption for ultra-large-scale images has not yet been effectively resolved.

[0006] Hybrid registration methods balance registration performance by integrating the advantages of multiple technologies. Multi-scale registration frameworks employ a hierarchical processing strategy, first completing initial registration at low-resolution levels, and then progressively refining to achieve accurate alignment of high-resolution images. This effectively balances registration efficiency and accuracy, but suffers from drawbacks such as complex multi-scale parameter optimization and the potential accumulation of local registration errors at each level. Registration methods based on imaging geometry constraints reduce parameter degrees of freedom by introducing the range-Doppler equation or RPC model, significantly improving the global consistency of large-area image registration. However, they are highly dependent on sensor parameter accuracy, and in practical applications, model errors may lead to a decrease in registration accuracy. These methods demonstrate significant theoretical potential, but practical deployment faces challenges related to algorithmic complexity and engineering implementation.

[0007] In view of this, the present invention is hereby proposed. Summary of the Invention

[0008] To address the aforementioned technical problems in existing technologies, this invention provides a method and system for large-scale SAR image registration based on irregular regional divisions, which solves the problem of non-uniform distribution of geometric distortion caused by complex terrain and varied ground features in existing large-area SAR image registration technologies.

[0009] To achieve the above objectives, the technical solution of the present invention is as follows: Firstly, a method for registering large-scale SAR images based on irregularly divided regions includes: S1. Based on preset multimodal rules, superpixel segmentation is performed on two SAR images of the same area acquired at different times to obtain superpixel blocks; S2. Based on the region similarity of the superpixel blocks, dynamically merge adjacent superpixel blocks in the superpixel blocks to obtain a superpixel region; S3. Based on the deformation complexity and feature point distribution of the superpixel region, determine the polynomial order of the preset adaptive local offset model; S4. Based on the model parameters of the adaptive local offset model of the superpixel region, perform global fusion optimization and region merging optimization to obtain the optimized superpixel region; S5. Crop the common area of ​​the optimized superpixel region to obtain the registration result of the SAR image.

[0010] Furthermore, the superpixel segmentation includes: S11. Initialize seed points, determine the theoretical number of superpixels based on the theoretical size of each superpixel when it is initially uniformly distributed, and determine the cluster center on the SAR image based on the theoretical number of superpixels. S12. Based on the image characteristic parameters of the SAR image, determine the distance between each pixel in the SAR image and all cluster centers within a preset range adjacent to it, and assign the pixel to the nearest cluster center to form a superpixel; wherein, the image characteristic parameters include intensity distance, texture difference and spatial distance; S13. Update the position of the cluster centers through iterative clustering; S14. Merge fragment regions with an area smaller than the preset standard into adjacent superpixels to form superpixel blocks.

[0011] Furthermore, the method for calculating the region similarity of the superpixel blocks is as follows: A dual-threshold merging criterion is defined, considering both feature point density and intensity distribution similarity. An improved SAR-SIFT algorithm is used to extract feature points within each superpixel, and the feature density difference is calculated. :

[0012] in, and Let A be the number of feature points, C be the area of ​​the superpixel region, D be the superpixel region currently under consideration, and D be another superpixel region adjacent to C. The similarity of intensity histograms is measured using the Bach distance. :

[0013] in , This is a normalized intensity histogram; Set the merging threshold: τ = 0.1 × global average density, η = 0.8; when adjacent regions meet the following conditions... If so, then merge.

[0014] Furthermore, the dynamic merging of adjacent superpixel blocks in the superpixel block includes: Construct a region adjacency graph, where nodes are superpixel blocks, edges represent the adjacency relationships between adjacent superpixels, and weights are... and The combined score; The formula for calculating the merging priority of all adjacent region pairs is as follows:

[0015] Will Edges are inserted into the queue in descending order of priority. Edges are taken from the top of the queue. If the merging condition is still met, the regions are merged. The region adjacency graph and the queue are updated. The intensity histogram and feature density of the new merged region are updated by area-weighted average.

[0016] Furthermore, step S3 specifically includes: S31. Determine the complexity of the superpixel region by quantifying the degree of nonlinearity of deformation within the superpixel region; S32. Determine the model order of the adaptive local offset model for each superpixel region based on the complexity of the superpixel region. S33. Optimize the model order by using regularization constraint parameters; S34. By introducing boundary consistency constraints, the optimized adjacent superpixel regions of different orders can be smoothly transitioned.

[0017] Furthermore, determining the complexity of the superpixel region by quantifying the nonlinearity of deformation within the superpixel region includes: S311, Define feature point density :

[0018] in, Superpixel region The number of feature points that successfully matched within the inner range. The superpixel area; S312. Use a first-order polynomial for fitting and calculate the initial linear residuals. :

[0019] Among them, the larger the residual, the stronger the deformation nonlinearity; S313. Calculate the standard deviation of elevation within the superpixel region based on DEM (Digital Elevation Model) data. :

[0020] in, The area of ​​the superpixel region. This is the elevation value, specifically representing the elevation value of the i-th pixel within the superpixel region. This represents the average elevation within the superpixel region. It is the set of all pixels within the superpixel region; S314. Calculate the comprehensive complexity score based on the feature point density, the initial linear residual, and the elevation standard deviation:

[0021] Among them, α+β+γ=1, typical values ​​α=0.4, β=0.4, γ=0.2.

[0022] Furthermore, step S32 specifically includes: For low-complexity regions where the complexity is less than the lower limit of the preset complexity scoring threshold, a first-order linear model is used: ; For medium complexity regions where the complexity falls within a preset complexity scoring threshold, a second-order polynomial model is used: ; For high-complexity regions where the complexity exceeds the upper limit of the preset complexity scoring threshold, a third-order polynomial model is used: .

[0023] Furthermore, the global fusion optimization includes: A global energy function is constructed by establishing a graph model; wherein, the nodes of the graph model are each superpixel region and its corresponding model parameters of the adaptive local offset model, the edges of the graph are the connection relationships between adjacent superpixel regions, and the weights are the inconsistency energy at the boundaries of adjacent superpixel regions; the global energy function includes a data term and a smoothing term. The model parameters are globally fused and optimized using the alternating direction multiplier method to ensure that the model parameters satisfy the constraints of the data term and the smoothing term.

[0024] Furthermore, the region merging optimization includes: Label all superpixel neighbor pairs and calculate their model parameter differences and residual change rates; Adjacent regions are arranged in ascending order of model parameter differences to form a queue, and the most similar superpixel regions are merged first. The most similar region pairs are taken from the queue. If the merging condition is met, the most similar region pairs are merged to form a new superpixel region, and the model parameters of the new superpixel region are updated. Update the adjacency graph and queue until there are no more superpixel regions to merge.

[0025] Secondly, a large-scale SAR image registration system based on irregularly segmented regions includes: The superpixel segmentation unit is used to perform superpixel segmentation on two SAR images of the same area acquired at different times, based on preset multimodal rules, to obtain superpixel blocks; The dynamic merging unit is used to dynamically merge adjacent superpixel blocks in the superpixel block according to the region similarity of the superpixel block to obtain a superpixel region; The model order determination unit is used to determine the polynomial order of the preset adaptive local offset model based on the deformation complexity and feature point distribution of each superpixel region. The region optimization unit is used to perform global fusion optimization and region merging optimization based on the model parameters of the adaptive local offset model of each superpixel region to obtain the optimized superpixel region. The cropping unit is used to crop the common area of ​​the optimized superpixel region to obtain the registration result of the SAR image.

[0026] Compared with existing technologies, the present invention provides a method and system for large-scale SAR image registration based on irregular regional segmentation. The method includes: segmenting two SAR images of the same area at different times into superpixel blocks based on preset multimodal rules; dynamically merging adjacent blocks according to the regional similarity of the superpixel blocks to obtain superpixel regions; determining the polynomial order of the adaptive local migration model based on the deformation complexity and feature point distribution of the superpixel regions; performing global fusion optimization and regional merging optimization based on model parameters to obtain optimized superpixel regions; and cropping the common areas of the optimized regions to obtain the registration result. The system includes: superpixel segmentation, dynamic merging, model order determination, regional optimization, and cropping units. The present invention takes into account both the adaptability to local distortion of complex large-area SAR images and the consistency of global registration, and is especially suitable for scenes with severe geometric deformation such as mountains and coastlines. Attached Figure Description

[0027] Figure 1 A flowchart of a large-scale SAR image registration method provided in an embodiment of the present invention; Figure 2 A flowchart of dynamic irregular partitioning of SAR images provided in an embodiment of the present invention; Figure 3 A flowchart of adaptive offset estimation and merging provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0028] The technical solution of the present invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are not all embodiments of the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0029] It should be noted that, unless otherwise specifically stated, the relative arrangement and numerical expressions of the components and steps described in these embodiments should not be construed as limiting the scope of the invention.

[0030] The following description of exemplary embodiments is merely illustrative and is not intended to limit the invention or its application or use in any way. Techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail herein, but where applicable, such techniques, methods, and apparatus should be considered part of this specification.

[0031] Terminology Explanation: The SAR-SIFT (Synthetic Aperture Radar Scale-Invariant Feature Transform) algorithm is a feature extraction and matching algorithm specifically designed for synthetic aperture radar (SAR) images. It improves the ability to describe edge structures and reduces the impact of random phase errors and intensity fluctuations commonly found in SAR images by replacing the original difference operation with an exponentially weighted mean ratio operator during the scale space construction process.

[0032] Digital Elevation Model (DEM) is an important source of data for studying and analyzing topography, watersheds, and feature identification. Because DEM data can reflect local topographic features at a certain resolution, a large amount of surface morphology information can be extracted from it. It can be used to draw contour lines, slope maps, aspect maps, perspective views, and 3D landscape maps, and can also be applied to the production of orthophotos, 3D topographic models, and map revision.

[0033] Example 1 See Figures 1-3 The flowchart of the large-scale SAR image registration method based on irregular regional division proposed in this invention includes the following steps: S1. Based on preset multimodal rules, superpixel segmentation is performed on two pre-acquired SAR images of the same area at different times to obtain superpixel blocks; specifically including: S11. Initialize seed points, determine the theoretical number of superpixels based on the theoretical size of each superpixel when initially uniformly distributed, and determine the cluster centers on the SAR image based on the theoretical number of superpixels.

[0034] In the process of initializing the seed point, the expected side length of the superpixel can be defined in advance. (Unit: pixels), representing the theoretical size of each superpixel when initially uniformly distributed. For high-resolution images =16~32, capturing fine structures, suitable for medium and low resolution images. =32~64, to improve computational efficiency, usually taken as... =32, balancing versatility and efficiency.

[0035] Based on the total number of pixels in the SAR image (W and H represent the range and azimuth pixel counts of the SAR image, respectively) and target superpixel area. The theoretical number of superpixels is calculated using the following formula:

[0036] Introducing the SAR image aspect ratio R=W / H, optimizing the superpixel grid distribution, and increasing the number of grids in the width direction. Number of grid cells in the height direction , Finally, constrain the minimum number of superpixels. To avoid segmentation failure due to excessively small image size, the maximum number of superpixels is constrained. To prevent superpixels from becoming too small.

[0037] After initializing the seed points, K initial cluster centers are uniformly distributed on the SAR image, with a spacing of [missing information]. , The total number of pixels. To avoid seed points falling in high gradient regions (such as edges), it is also necessary to calculate the gradient for a 3×3 window around each seed point, and move the initial cluster center to the position of minimum gradient as the cluster center for superpixel segmentation (hereinafter collectively referred to as cluster center).

[0038] S12. Based on the image characteristic parameters of the SAR image, determine the distance between each pixel in the SAR image and all cluster centers within a preset range adjacent to it, and assign the pixel to the nearest cluster center to form a superpixel. The image characteristic parameters include intensity distance, texture difference, and spatial distance.

[0039] The purpose of this step is to design a multimodal distance metric that integrates intensity, texture, and spatial information D.

[0040]

[0041] in, Indicates intensity distance. Indicates texture differences, The spatial distance is represented by α, β, and γ, which are the weights of the three image feature parameters, respectively.

[0042] Specifically, for illustration, intensity distance Normalized variance using a local window (5×5) is employed to suppress the influence of speckle noise.

[0043] in, The intensity value of the i-th pixel within the window. The average intensity within the window. This represents the total number of pixels in the SAR image. This represents the number of range pixels in the SAR image. Calculate contrast It reflects the sharpness of local edges and is sensitive to terrain undulations; it calculates correlation. It describes the consistency of texture direction and distinguishes areas such as farmland and forest.

[0044] in, Represents pixel value, This represents the average value of all pixels.

[0045] in, and These represent pixel values ​​at two different pixel locations or in different wavelength bands. and Each is its own average value; The standard deviation of GLCM features within the calculation window is used as texture difference. The specific formula is as follows:

[0046] Calculate spatial distance The specific formula is as follows:

[0047] in,( () represents any pixel coordinate, ) represents the current superpixel center coordinates.

[0048] Typical weight values ​​for optimization on the validation set using grid search: It emphasizes the dominant role of strength and texture.

[0049] S13. Update the position of the cluster center through iterative clustering.

[0050] Specifically, as an example, for each pixel, calculate the distance D between it and all cluster centers within a 2S×2S range (where S represents the number of pixels and 2S×2S represents the search range, i.e., each pixel will consider the cluster centers within a 2S×2S range around it), assign it to the nearest center, then update the cluster center position to the average coordinates of the pixel to which it belongs, and recalculate the intensity and texture features.

[0051] In a specific embodiment of the present invention, the termination condition for the iterative clustering update is: 10 iterations or the center movement distance is less than 1 pixel.

[0052] S14. Merge fragment regions with an area smaller than the preset standard into adjacent superpixels to form superpixel blocks.

[0053] This step involves forcing connectivity for superpixels; specifically, fragmented regions with an area smaller than 0.5S² can be removed and merged into adjacent superpixels to form superpixel blocks.

[0054] S2. Based on the regional similarity of the superpixel blocks, dynamically merge adjacent superpixel blocks to obtain a superpixel region; the main task is to merge adjacent superpixel blocks to form irregular regions that adapt to the feature distribution and avoid over-segmentation. Specifically, this includes: S21. Region similarity measurement: The calculation method for the region similarity of superpixel blocks is as follows: A dual-threshold merging criterion is defined, considering both feature point density and intensity distribution similarity. An improved SAR-SIFT algorithm is used to extract feature points within each superpixel, and the feature density difference is calculated. :

[0055] in, and Let A be the number of feature points, C be the area of ​​the superpixel region, D be the superpixel region currently under consideration, and D be another superpixel region adjacent to C. The similarity of intensity histograms is measured using the Bach distance. :

[0056] in , This is a normalized intensity histogram, where B=16 is the number of bins; Set the merge threshold: , When adjacent regions satisfy If so, then merge.

[0057] S22. Dynamically merging adjacent superpixel blocks in the superpixel block specifically includes: Construct a region adjacency graph, where nodes are superpixel blocks, edges represent the adjacency relationships between adjacent superpixels, and weights are Δfeat and Δfeat. The combined score; Calculate the merging priority for all adjacent region pairs:

[0058] Will Edges are inserted into the queue in descending order of priority. Edges are taken from the top of the queue. If the merging condition is still met, the regions are merged. The region adjacency graph and the queue are updated. The intensity histogram and feature density of the new merged region are updated by area-weighted average.

[0059] In addition, in this embodiment, after the region merging, regions with an area greater than 4S² (where S represents the number of pixels; 4S² represents the area of ​​a region, specifically 4 times the square of the number of pixels) are forcibly segmented to prevent a single region from covering too many heterogeneous objects; furthermore, morphological closing operations (3×3 cross kernels) can be used to smooth the boundaries of irregular regions to reduce the jagged effect of superpixel region boundaries.

[0060] After obtaining the final superpixel region, step S3 can be performed.

[0061] S3. Based on the deformation complexity and feature point distribution of the superpixel region, determine the pre-defined polynomial order of the adaptive local migration model. This mainly involves dynamically selecting the optimal polynomial order (e.g., 1st, 2nd, or 3rd order) for each superpixel region to balance the accuracy of the adaptive local migration model with the risk of overfitting. The selection of the polynomial order is determined based on the deformation complexity and feature point distribution of each superpixel region. Specific steps include region complexity evaluation, adaptive order selection, regularization parameter optimization, and cross-model smoothing constraints. Specifically, this includes: S31. Determine the complexity of the superpixel region by quantizing the degree of nonlinearity of deformation within the superpixel region. Quantifying the degree of nonlinearity of deformation within the region provides a basis for order selection. Specifically, this includes the following steps: S311, Define feature point density The specific formula is as follows:

[0062] in, Superpixel region The number of feature points that successfully matched within the inner range. The area represents the superpixel region; higher density may require a higher-order model. S312. Use a first-order polynomial for fitting and calculate the initial linear residuals. The specific formula is as follows:

[0063] The larger the residual, the stronger the nonlinearity of the deformation.

[0064] S313. Calculate the standard deviation of elevation within the superpixel region based on DEM (Digital Elevation Model) data. The specific formula is as follows:

[0065] in, The area of ​​the superpixel region. This is the elevation value, specifically representing the elevation value of the i-th pixel within the superpixel region. This represents the average elevation within the superpixel region. It is the set of all pixels within a superpixel region; regions with large terrain undulations are more likely to require a higher-order model.

[0066] S314: The comprehensive complexity score is calculated based on the feature point density, the initial linear residual, and the elevation standard deviation. The specific formula is as follows:

[0067] in, Typical value .

[0068] S32. Determine the model order of the adaptive local offset model for each superpixel region based on the complexity of the superpixel region. This step is used to select the adaptive polynomial order and divide the superpixel region into low-complexity region, medium-complexity region and high-complexity region according to the preset complexity scoring threshold.

[0069] In this embodiment, the preset complexity scoring threshold is set to [0.3, 0.6]. Based on the complexity of the superpixel region, the model order of the adaptive local offset model for each superpixel region is determined, specifically including: For complexity less than the lower limit of the preset complexity scoring threshold ( For the low-complexity region <0.3), a first-order linear model (affine transformation) is used: ; For complexity within the preset complexity scoring threshold (0.3≤ For medium complexity regions (<0.6), a second-order polynomial model is used: ; For complexity values ​​exceeding the upper limit of the preset complexity scoring threshold For high-complexity regions (≥0.6), a third-order polynomial model is used:

[0070] Among them, if the standard deviation of the residuals after fitting a certain order model Still above the threshold Then increase the order until the following condition is met: if Or it may reach the maximum order. The threshold τ is set according to the image resolution, usually 1~2 pixels.

[0071] S33. Optimize the model order by using regularization constraint parameters; Higher-order models are prone to overfitting and require regularization constraints on their parameters. Therefore, a regularization design is implemented to penalize higher-order coefficients and prevent excessive fluctuations. The specific formula is as follows:

[0072] Only penalty in the second-order model The third-order model penalizes the coefficients of all third-order terms, and the weights... Adaptive, the specific formula is:

[0073] in, Based on the weights, For regions with high complexity and few feature points, stronger regularization is used to minimize the number of effective points.

[0074] S34. By introducing boundary consistency constraints, a smooth transition between models of different orders in adjacent superpixel regions after optimization is achieved. To ensure a smooth transition between models of different orders in adjacent regions, in a specific embodiment of the present invention, boundary consistency constraints are introduced for adjacent regions and By sampling several points on their common boundary, the offsets predicted by the two models are forced to be continuous. The specific formula is as follows:

[0075] This constraint is added as a soft condition to the objective function to obtain the global optimization function:

[0076] Where μ is the smoothing weight, usually taken as μ=10. .

[0077] S4. Based on the model parameters of the adaptive local offset model of the superpixel region, perform global fusion optimization and region merging optimization to obtain the optimized superpixel region; specifically including: S41. Global fusion optimization includes: A global energy function is constructed by establishing a graph model; wherein, the nodes of the graph model are each superpixel region and its corresponding model parameters of the adaptive local offset model, the edges of the graph are the connection relationships between adjacent superpixel regions, and the weights are the inconsistency energy at the boundaries of adjacent superpixel regions; the global energy function includes a data term and a smoothing term. The model parameters are globally fused and optimized using the alternating direction multiplier method to ensure that the model parameters satisfy the constraints of the data term and the smoothing term.

[0078] More specifically, as an example, we can build a graph model where each node represents a region. and its corresponding local offset model parameters The edges of the graph represent the connectivity between adjacent regions, and the weights are the inconsistencies at the boundaries of adjacent superpixel regions. A global energy function is constructed based on this, and the total energy is minimized by optimizing the model parameters.

[0079] The global energy function includes a data term and a smoothing term: ; Among them, data items Measuring the matching accuracy of a local model within its own region: ; This represents the offset predicted by the current model. This represents a regularization term to prevent overfitting.

[0080] Smoothing Term Force adjacent regions to have the same offset at the boundary: ; Then, the alternating direction multiplier method is used for optimization at each boundary point. Calculate the prediction offset difference between adjacent models, generate consistency constraints, and use the least squares method to jointly optimize all model parameters to satisfy the data term and smoothing term constraints: ,in For Jacobian matrices, To observe the offset, repeat the above steps until the energy change is less than the threshold or the maximum number of iterations is reached.

[0081] S42. Regional merging optimization includes: Label all superpixel neighbor pairs and calculate their model parameter differences and residual change rates; Adjacent regions are arranged in ascending order of model parameter differences to form a queue, and the most similar superpixel regions are merged first. The most similar region pairs are taken from the queue. If the merging condition is met, the most similar region pairs are merged to form a new superpixel region, and the model parameters of the new superpixel region are updated. Update the adjacency graph and queue until there are no more superpixel regions to merge. Merging adjacent superpixel regions requires the following conditions to be met: (1) Model similarity: the difference in offset model parameters between two superpixel regions is less than the threshold. (2) Consistency of features: the feature points in the merged region are evenly distributed and there is no significant heterogeneity; (3) Deformation continuity: The residuals of the merged single model in the original two regions do not increase significantly.

[0082] Therefore, before merging adjacent superpixels, the parameter differences are first defined: Define the rate of change of residuals: .

[0083] Then, all adjacent region pairs are labeled, their model parameter differences and residual change rates are calculated, adjacent region pairs are sorted in ascending order of parameter differences, and the most similar regions are merged first. The most similar region pairs are then removed from the queue. If the merger conditions are met, then merge into Update its model parameters:

[0084] in, Update the adjacency graph and queue to the original number of pixels until there are no more superpixel regions to merge.

[0085] S5. Crop the common area of ​​the optimized superpixel region to obtain the registration result of the SAR image.

[0086] In summary, this invention addresses the problem of non-uniform geometric distortion caused by complex terrain and varied ground features in large-area SAR image registration. It achieves efficient and accurate registration through a three-step core design: First, based on local features such as image space, intensity, and texture, the large area is adaptively segmented into multiple differentiated sub-regions, avoiding the disruption of ground feature boundaries caused by traditional regular block division. Second, by dynamically selecting the polynomial order and combining it with multi-model collaborative optimization, the shortcomings of traditional fixed-order models—sensitivity to high-order noise or insufficient low-order representation—are addressed. Finally, a global fusion and region merging optimization framework is constructed, achieving smooth fusion of region registration results by minimizing the global energy function. This method effectively balances the adaptability to local distortion in complex large-area SAR images with global registration consistency, and is particularly suitable for scenes with severe and non-uniform geometric deformation, such as mountains and coastlines, providing a reliable solution for large-scale SAR image registration.

[0087] Example 2 This invention proposes a large-scale SAR image registration system based on irregular regional division, comprising: D1, the superpixel segmentation unit, is used to perform superpixel segmentation on two SAR images of the same area at different times, based on preset multimodal rules, to obtain superpixel blocks; D2, Dynamic merging unit, used to dynamically merge adjacent superpixel blocks in the superpixel block according to the region similarity of the superpixel block to obtain a superpixel region; D3, Model Order Determination Unit, is used to determine the polynomial order of the preset adaptive local offset model based on the deformation complexity and feature point distribution of each superpixel region. D4, the region optimization unit, is used to perform global fusion optimization and region merging optimization based on the model parameters of the adaptive local offset model of each superpixel region to obtain the optimized superpixel region; D5, the cropping unit, is used to crop the common area of ​​the optimized superpixel region to obtain the registration result of the SAR image.

[0088] Example 3 See Figure 4 , Figure 4 The electronic device proposed in this invention, which is equipped with a large-scale SAR image registration system based on irregular regional division, specifically includes: At least one processor; and, A memory that is communicatively connected to at least one processor; wherein, The memory stores a computer program that can be executed by at least one processor to enable the at least one processor to perform the steps in the large-scale SAR image registration method based on irregular regional division in Embodiment 1.

[0089] Those skilled in the art will understand that the structure shown in the figures does not constitute a limitation on the electronic device 1, and may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0090] For example, although not shown, the electronic device 1 may also include a power supply (such as a battery) to power the various components. Preferably, the power supply can be logically connected to the at least one processor 10 through a power management device, thereby enabling functions such as charging management, discharging management, and power consumption management. The power supply may also include one or more DC or AC power supplies, recharging devices, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components. The electronic device 1 may also include various sensors, Bluetooth modules, Wi-Fi modules, etc., which will not be described in detail here.

[0091] Furthermore, the electronic device 1 may also include a network interface. Optionally, the network interface may include a wired interface and / or a wireless interface (such as a Wi-Fi interface, a Bluetooth interface, etc.), which is typically used to establish communication connections between the electronic device 1 and other electronic devices.

[0092] Optionally, the electronic device 1 may further include a user interface, which may be a display, an input unit (such as a keyboard), or a standard wired or wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen, etc. The display may also be appropriately referred to as a screen or display unit, used to display information processed in the electronic device 1 and to display a visual user interface.

[0093] It should be understood that the embodiments described are for illustrative purposes only and are not limited to this structure in the scope of the patent application.

[0094] The large-scale SAR image registration program 12 based on irregular regional division stored in the memory 11 of the electronic device 1 is a combination of multiple instructions. When run in the processor 10, it can implement the following steps: S100: Based on preset multimodal rules, superpixel segmentation is performed on two SAR images of the same area acquired at different times to obtain superpixel blocks; S200: Based on the region similarity of the superpixel blocks, dynamically merge adjacent superpixel blocks in the superpixel blocks to obtain a superpixel region; S300: Determine the polynomial order of the preset adaptive local offset model based on the deformation complexity and feature point distribution of each superpixel region; S400: Perform global fusion optimization and region merging optimization based on the model parameters of the adaptive local offset model for each superpixel region to obtain the optimized superpixel region; S500: Crop the common area of ​​the optimized superpixel region to obtain the registration result of the SAR image.

[0095] Specifically, the specific implementation method of the processor 10 for the above instructions can be referred to the description of the corresponding embodiment in the large-scale SAR image registration method based on irregular regional division, and will not be repeated here.

[0096] Furthermore, if the modules / units integrated in the electronic device 1 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. The computer-readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, or a read-only memory (ROM).

[0097] The large-scale SAR image registration method and system based on irregular regional division proposed according to the present invention have been described above by way of example with reference to the accompanying drawings. However, those skilled in the art should understand that various modifications can be made to the large-scale SAR image registration method and system based on irregular regional division proposed by the present invention without departing from the scope of the present invention. Therefore, the scope of protection of the present invention should be determined by the content of the appended claims.

[0098] In summary, the present invention has the following advantages: 1. A dynamic and irregular regional registration strategy is proposed to automatically segment large-area SAR images into multiple differentiated small regions, replacing the traditional fixed block mode, which has stronger adaptability. 2. It is suitable for scenarios with dramatic terrain undulations and diverse land features, breaking through the registration limitations of traditional models in complex areas; 3. Global fusion and regional merging optimization method: After estimating the local migration model (LPM) of each region, construct the graph energy model minimization constraint, solve the global energy function, minimize the total energy by optimizing the model parameters, merge the registration results of different regions, eliminate the contradiction between adjacent regional models, and ensure a smooth transition of registration results.

[0099] The above specific embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for large-scale SAR image registration based on irregularly divided regions, characterized in that, include: S1. Based on preset multimodal rules, superpixel segmentation is performed on two SAR images of the same area acquired at different times to obtain superpixel blocks; S2. Based on the region similarity of the superpixel blocks, dynamically merge adjacent superpixel blocks in the superpixel blocks to obtain a superpixel region; S3. Based on the deformation complexity and feature point distribution of the superpixel region, determine the polynomial order of the preset adaptive local offset model; Step S3 specifically includes: S31. Determine the complexity of the superpixel region by quantifying the degree of nonlinearity of deformation within the superpixel region; including: S311, Define feature point density : in, Superpixel region The number of feature points that successfully matched within the inner range. The superpixel area; S312. Use a first-order polynomial for fitting and calculate the initial linear residuals. : Among them, the larger the residual, the stronger the deformation nonlinearity; S313. Calculate the standard deviation of elevation within the superpixel region based on the DEM data. : in, The area of ​​the superpixel region. This is the elevation value, specifically representing the elevation value of the i-th pixel within the superpixel region. This represents the average elevation within the superpixel region. It is the set of all pixels within the superpixel region; S314. Calculate the comprehensive complexity score based on the feature point density, initial linear residual, and elevation standard deviation: Among them, α+β+γ=1, typical values ​​α=0.4, β=0.4, γ=0.2; S32. Determine the model order of the adaptive local offset model for each superpixel region based on the complexity of the superpixel region; specifically including: For low-complexity regions where the complexity is less than the lower limit of the preset complexity scoring threshold, a first-order linear model is used: ; For medium complexity regions where the complexity falls within a preset complexity scoring threshold, a second-order polynomial model is used: ; For high-complexity regions where the complexity exceeds the upper limit of the preset complexity scoring threshold, a third-order polynomial model is used: ; S33. Optimize the model order by regularizing the constraint parameters; S34. By introducing boundary consistency constraints, the optimized adjacent superpixel regions of different orders of models can be smoothly transitioned. S4. Based on the model parameters of the adaptive local offset model of the superpixel region, perform global fusion optimization and region merging optimization to obtain the optimized superpixel region; S5. Crop the common area of ​​the optimized superpixel region to obtain the registration result of the SAR image.

2. The method for large-scale SAR image registration based on irregular regional division according to claim 1, characterized in that, The superpixel segmentation includes: S11. Initialize seed points, determine the theoretical number of superpixels based on the theoretical size of each superpixel when it is initially uniformly distributed, and determine the cluster centers on the SAR image based on the theoretical number of superpixels. S12. Based on the image characteristic parameters of the SAR image, determine the distance between each pixel in the SAR image and all cluster centers within a preset range, and assign the pixel to the nearest cluster center to form a superpixel; wherein, the image characteristic parameters include intensity distance, texture difference and spatial distance; S13. Update the position of the cluster center through iterative clustering; S14. Merge fragment regions with an area smaller than the preset standard into adjacent superpixels to form superpixel blocks.

3. The method for large-scale SAR image registration based on irregular regional division according to claim 2, characterized in that, The method for calculating the region similarity of the superpixel blocks is as follows: A dual-threshold merging criterion is defined, considering both feature point density and intensity distribution similarity. An improved SAR-SIFT algorithm is used to extract feature points within each superpixel, and the feature density difference is calculated. : in, and Let A be the number of feature points, C be the area of ​​the superpixel region, D be the superpixel region currently being considered, and D be another superpixel region adjacent to C. The similarity of intensity histograms is measured using the Bach distance. : in , This is a normalized intensity histogram; Set the merging threshold: τ = 0.1 × global average density, η = 0.8; when adjacent regions meet the following conditions... If so, then merge.

4. The method for large-scale SAR image registration based on irregular regional division according to claim 3, characterized in that, The adjacent superpixel blocks in the dynamically merged superpixel blocks include: Construct a region adjacency graph, where nodes are superpixel blocks, edges represent the adjacency relationships between adjacent superpixels, and weights are... and The combined score; The formula for calculating the merging priority of all adjacent region pairs is as follows: Will Edges are inserted into the queue in descending order of priority. Edges are taken from the top of the queue. If the merging condition is still met, the regions are merged. The region adjacency graph and the queue are updated. The intensity histogram and feature density of the new merged region are updated by area-weighted average.

5. The method for large-scale SAR image registration based on irregular regional division according to claim 1, characterized in that, The global fusion optimization includes: A global energy function is constructed by establishing a graph model; wherein, the nodes of the graph model are each superpixel region and its corresponding model parameters of the adaptive local offset model, the edges of the graph are the connection relationships between adjacent superpixel regions, and the weights are the inconsistency energy at the boundaries of adjacent superpixel regions; the global energy function includes a data term and a smoothing term. The model parameters are globally fused and optimized using the alternating direction multiplier method to ensure that the model parameters satisfy the constraints of the data term and the smoothing term.

6. The method for large-scale SAR image registration based on irregular regional division according to claim 1, characterized in that, The region merging optimization includes: Label all superpixel neighbor pairs and calculate their model parameter differences and residual change rates; Adjacent regions are arranged in ascending order of model parameter differences to form a queue, and the most similar superpixel regions are merged first. Among them, the most similar region pairs are taken from the queue. If the merging condition is met, the most similar region pairs are merged to become a new superpixel region, and the model parameters of the new superpixel region are updated. Update the adjacency graph and queue until there are no more superpixel regions to merge.

7. A large-scale SAR image registration system based on irregularly divided regions, characterized in that, The method for large-scale SAR image registration based on irregular regional division as described in any one of claims 1-6 includes: The superpixel segmentation unit is used to perform superpixel segmentation on two SAR images of the same area acquired at different times based on preset multimodal rules, to obtain superpixel blocks; A dynamic merging unit is used to dynamically merge adjacent superpixel blocks in the superpixel block based on the region similarity of the superpixel block to obtain a superpixel region; The model order determination unit is used to determine the polynomial order of the preset adaptive local offset model based on the deformation complexity and feature point distribution of each superpixel region. The region optimization unit is used to perform global fusion optimization and region merging optimization based on the model parameters of the adaptive local offset model of each superpixel region to obtain the optimized superpixel region. The cropping unit is used to crop the common area of ​​the optimized superpixel region to obtain the registration result of the SAR image.

Images