Remote sensing image registration method based on hierarchical recursive coarse registration and physical scale alignment

Abstract

This invention relates to the fields of remote sensing technology, computer vision, and artificial intelligence, and discloses a remote sensing image registration method based on hierarchical recursive coarse matching and physical scale alignment. The method includes: reading geographic metadata and Earth affine transformation parameters of multi-source high-resolution reference images and images to be registered; performing logical segmentation and normalized scaling feature extraction using a hierarchical recursive grid generation module, and obtaining initial homography transformation parameters through coordinate inverse kinematics and random sampling consistency algorithms; after preliminary correction of the images to be registered, calculating the relative scaling factor and constructing a dynamic field-of-view feature pyramid based on the field-of-view magnification set, and performing forced unified resampling alignment; scheduling a multi-threaded distributed concurrent engine to load multi-scale tiles, capturing local temporary index information through a model shell interceptor, and reconstructing using a full-batch global identifier matrix. This invention eliminates macroscopic geometric distortion, overcomes scale variance, avoids tensor coordinate contamination, and improves registration efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing technology, computer vision and artificial intelligence, and in particular to a remote sensing image registration method based on hierarchical recursive coarse matching and physical scale alignment. Background Technology

[0002] In modern high-resolution remote sensing, UAV photogrammetry, and multi-source image fusion, dense matching of image features across resolutions and large field-of-view variations is a prerequisite for achieving accurate geospatial stitching and deformation correction. Traditional feature matching methods (such as SIFT and SURF) follow a sequential logic of first detecting feature points and then extracting descriptors, heavily relying on local pixel gradient changes. When dealing with highly repetitive texture features (such as agricultural greenhouses and forest vegetation) or weakly textured areas (such as water surfaces, deserts, and snow cover), matching often fails because it cannot extract feature points with significant discriminative power. In recent years, detector-free dense matching networks based on deep learning (such as Transformer-based models) have improved the robustness of feature matching in complex scenes through global attention mechanisms.

[0003] When dealing with ultra-large-scale remote sensing image matching tasks, existing deep learning matching techniques and engineering solutions do not fully integrate the physical spatial characteristics of remote sensing data, and still have the following shortcomings:

[0004] First, the fixed-step sliding window strategy is prone to computational redundancy and attention mechanism failure under large field-of-view (FoV) offsets. Current large-image processing often employs a fixed-step sliding window segmentation strategy. When multi-source remote sensing images lack accurate geographic priors or have excessively large initial spatial offsets, this strategy generates a large number of invalid tiles without shared viewing areas (such as data-free areas or non-overlapping edges). Because the network model is limited by a fixed input size, the lack of shared features within local tiles prevents the cross-image attention mechanism from establishing effective feature associations, thus causing local region matching failure.

[0005] Secondly, differences in physical ground sampling distance (GSD) lead to spatial frequency mismatches and abnormal feature associations. Multi-source sensor images often exhibit significant differences in absolute physical resolution (GSD). Existing deep learning models typically rely on internal network modules (such as Feature Pyramid Networks (FPN) and pooling layers) to implicitly learn scale invariance. However, when the physical resolution differences of the input images are too large, the feature descriptors extracted from a fixed pixel-level receptive field will contain inconsistent spatial frequency information. This method, which relies solely on network parameters to fit the absolute physical scale differences, has low robustness and is prone to causing the model to incorrectly associate structural and texture features at different scales.

[0006] Third, the limitations of local tensor indexing in distributed concurrent environments lead to global coordinate mapping errors. In large-scale data processing, image slices are typically loaded concurrently in batches within multi-process or multi-GPU environments. Existing deep learning frameworks only allocate local temporary indexes for slices within a batch. After the model outputs feature coordinates through forward inference, these local indexes lack a correspondence with the absolute spatial location of the entire image. Under asynchronous concurrent scheduling, the system struggles to accurately reverse-map local tensor coordinates to the global geographic coordinate system, easily causing coordinate matching errors across images and batches.

[0007] The invention patent CN105809693A proposes an image registration method based on deep neural networks, which performs global and local bidirectional matching through feature point extraction. However, this method typically relies on static image cropping when dealing with significant initial viewpoint deviations, failing to establish a dynamic mesh generation mechanism based on constraints from an absolute geographic coordinate system (such as geographic affine transformation parameters). Therefore, this method has limited accuracy in extracting spatially valid regions of intersection (ROIs), is prone to generating unnecessary computational loads, and struggles to effectively correct macroscopic geometric distortions at extremely low resolutions.

[0008] Patent CN113283525B discloses a deep learning-based image matching method that processes features of different resolutions through feature extraction and fusion modules to obtain a high-resolution fused feature map. This type of technology primarily relies on hidden layers of neural networks to handle resolution differences, without introducing a dynamic reading and resampling alignment mechanism based on real-world physical GSD during the data loading and preprocessing stages. When the input is a heterogeneous tensor without a unified physical benchmark, the model is susceptible to resolution mismatch in cross-sensor matching tasks, leading to a decrease in the matching accuracy of feature regions.

[0009] Publication number CN109711444A proposes a method for image geometric correction using depth features. Such methods are often based on single-image or serial processing flows, lacking batch-level tensor identification management mechanisms in distributed parallel processing architectures. Because they fail to establish and pass a global grid index during the data loading phase, and do not replace the original local temporary index at the model output, these techniques struggle to guarantee the accuracy of local feature coordinate transformation to the global absolute coordinate system and data consistency when performing large-scale image segmentation and concurrent processing.

[0010] In summary, existing technical solutions have several problems when dealing with dense matching of large-scale multi-source remote sensing images. These problems include low efficiency in initial offset processing of large-scale fields of view, difficulty in overcoming the scale variance trap caused by physical ground sampling distance (GSD), and the tendency to cause global coordinate system collapse under distributed concurrent architecture. Summary of the Invention

[0011] The purpose of this invention is to provide a remote sensing image registration method based on hierarchical recursive coarse matching and physical scale alignment. This method solves the technical problems of existing remote sensing image matching techniques, such as low efficiency in processing large-scale initial field of view offsets, difficulty in overcoming the scale variance trap caused by physical ground sampling distance, and the tendency to generate tensor local index aliasing and global geographic coordinate system collapse under distributed concurrent architecture.

[0012] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0013] The remote sensing image registration method based on hierarchical recursive coarse matching and physical scale alignment includes the following steps:

[0014] Step 1: Read the geographic metadata and Earth affine transformation parameters of the multi-source high-resolution reference image and the image to be registered;

[0015] Step 2: Using the hierarchical recursive mesh generation module, a multi-level mesh is constructed to logically segment and extract normalized scaling features from the multi-source high-resolution reference image and the image to be registered. The initial homography transformation parameters for eliminating the large initial field of view offset are obtained by cross-level aggregation calculation using coordinate inverse solution and random sampling consistency algorithm.

[0016] Step 3: Using the multi-source high-resolution reference image as a reference, after performing preliminary correction on the image to be registered according to the initial homography transformation parameters, calculate the relative scaling factor based on the physical resolution of the multi-source high-resolution reference image and the physical resolution of the image to be registered. Combined with the preset field of view magnification set, construct a dynamic field of view feature pyramid in the underlying data pipeline, and perform forced unified resampling alignment on the multi-source high-resolution reference image data matrix and the image to be registered data matrix.

[0017] Step 4: For the images that have completed preliminary correction and unified resampling alignment, schedule a multi-threaded distributed concurrent engine to load multi-scale slicing tasks. Capture local temporary index information of the feature queue output by the forward inference of the deep learning matching model through the model shell interceptor. Use the full batch global identifier matrix to perform addressing reconstruction to replace the local temporary index information. Use the bounding box topological inclusion relationship to perform cross-validation to filter out anomalies and generate a dense feature point set with unique absolute global coordinates across processes.

[0018] The global spatial transformation matrix is ​​calculated based on a dense feature point set. Global pixel resampling is then performed on the image to be registered to output a high-resolution aligned remote sensing image.

[0019] As a preferred embodiment of the present invention, in step one, the input multi-source high-resolution reference image and the image to be registered are analyzed using an affine transformation model of the Earth. The geographic metadata six-parameter matrix is ​​represented as follows: Arbitrary pixel space coordinates Projected onto real geophysical space coordinates The mathematical formula is:

[0020]

[0021] in, Represents the absolute geographic spatial x-coordinate, Represents the absolute geographic spatial vertical coordinate. and The absolute geographic coordinates representing the origin at the top left corner of the remote sensing image. and The horizontal and vertical ground sampling distances of remotely sensed images, i.e., pixel spatial resolution. and Characterizing the rotational skew of remotely sensed images, Represents the x-coordinate of pixels. Represents the vertical coordinate of pixels.

[0022] As a preferred technical solution of the present invention, in step two, the height and width dimensions of the target input image are set, and a hierarchical set composed of multi-level layered grids is constructed. Each level of layered grid has a corresponding number of horizontal grid segments and a number of vertical grid segments. For any level of grid division, the multi-source high-resolution reference image and the image to be registered are cut into sub-images multiplied by the number of horizontal segments and the number of vertical segments of the current level grid. The height pixel value of each sub-image is the floor value of the target input image height divided by the number of horizontal segments of the current level grid, and the width pixel value of each sub-image is the floor value of the target input image width divided by the number of vertical segments of the current level grid.

[0023] As a preferred technical solution of the present invention, in step two, the cropped sub-image is scaled to the standard normalized target size required by the deep learning matching model through a bilinear interpolation function to form a normalized target size image.

[0024] In step two, the normalized target size image is input into the deep learning feature extraction backbone network to extract a set of matching point pairs. Using the local grid horizontal offset and local grid vertical offset of each layer grid, as well as the corresponding scaling factor, the normalized horizontal and vertical coordinates are mapped back to the absolute pixel horizontal and vertical coordinates of the entire image of the multi-source high-resolution reference image or the image to be registered. The scaling factor is the ratio of the sub-image width pixel value to the standard normalized target size and the ratio of the sub-image height pixel value to the standard normalized target size.

[0025] Furthermore, in step two, all absolute pixel coordinate point pairs generated by the layered grid are aggregated, and the global homography matrix is ​​solved using the random sampling consensus algorithm. The global homography matrix satisfies the mapping relationship between the homogeneous expression of matching coordinates on the multi-source high-resolution reference image and the homogeneous expression of matching coordinates on the image to be registered. Abnormal coordinate points are removed by the random sampling consensus algorithm to obtain the initial homography transformation parameters.

[0026] As a preferred technical solution of the present invention, in step three, the physical resolution of the reference image of the multi-source high-resolution reference image and the physical resolution of the image to be registered are read, and the physical scale relative scaling factor is calculated. The value of the physical scale relative scaling factor is the larger value of the physical resolution of the reference image and the physical resolution of the image to be registered divided by the smaller value.

[0027] As a preferred embodiment of the present invention, in step three, if the physical resolution of the reference image is less than the physical resolution of the image to be registered, then the scaling factor of the multi-source high-resolution reference image is assigned equal to the physical scale relative scaling factor, and the scaling factor of the image to be registered is equal to 1.0; the standard input tensor size of the model is set. The formula for calculating the actual physical read window size dynamically extracted from the front-end data stream is:

[0028]

[0029] in, The actual physical readout window size represents the multi-source high-resolution reference image. Represents the actual physical readout window size of the image to be registered. The size of the standard input tensor of the representative model. The scaling factor representing the multi-source high-resolution reference image. The scaling factor represents the image to be registered. This represents the floor operation.

[0030] As a preferred technical solution of the present invention, in step three, in combination with a preset set of field of view magnification, the physical width of the real physical reading window is forcibly magnified by a corresponding factor in the underlying data pipeline to construct a dynamic field of view feature pyramid, and the input model is covered by a standardized tensor of standard size through a bilinear interpolation function; wherein, each field of view magnification in the preset set of field of view magnification is not greater than the physical scale relative scaling factor.

[0031] As a preferred embodiment of the present invention, in step four, when the multi-threaded distributed concurrent engine loads the multi-scale slicing task, a unique cross-process global grid identifier is assigned to each valid slice pushed into the queue, and a full-batch global identifier matrix is ​​constructed. ,in Represents batch size. Represents a set of integers.

[0032] Furthermore, in step four, after the data is fed into the deep learning matching model for forward inference, the tensor queue with local temporary index information output by the deep learning matching model is intercepted. ,in This represents the total number of feature points extracted in the current batch; a one-dimensional dimensionality reduction mapping is performed on the global identifier matrix of the entire batch mounted in the video memory to reconstruct it into absolute geographic identity labels. The mathematical formula is:

[0033]

[0034] in, Representing the The absolute geographic identity label corresponding to each feature point Represents the global identifier matrix for the entire batch. Representing the Each feature point contains local temporary index values. This represents the total number of feature points extracted in the current batch. It represents a logical universal quantifier, referring to any value within a range. The feature point number is denoted as .

[0035] As a preferred embodiment of the present invention, in step four, a batch consistency verification function is constructed for each extracted feature coordinate. By comparing the absolute pixel coordinates of the entire image after inverse kinematics with the absolute geographic identity tags The topological inclusion relationship of the bounding boxes is cross-validated. The topological inclusion relationship of the bounding boxes refers to the spatial inclusion relationship used to determine whether the feature coordinates are located within the boundary of the mesh bounding box. The real-time outlier detection and filtering formula is as follows:

[0036]

[0037] in, Representing the The validity determination result of each feature coordinate. Representing the Each feature coordinate, Represents absolute geographical identity tags The associated real geographic grid bounding box boundary function.

[0038] As a preferred technical solution of the present invention, in step five, the effective feature point set after bounding box cross-validation is used to solve the global spatial transformation matrix by the least squares method, and the global pixel resampling is performed on the image to be registered based on the global spatial transformation matrix to output a high-resolution aligned remote sensing image.

[0039] As a preferred technical solution of the present invention, when the physical extent of the real physical readout window is forcibly enlarged in the underlying data pipeline to construct a dynamic field-of-view feature pyramid, when the physical scale relative scaling factor is greater than a preset threshold of 1.2, an anti-aliasing discrete low-pass filter based on physical ground sampling distance adaptive is performed on the real physical readout window of the multi-source high-resolution reference image or the real physical readout window of the image to be registered. The standard deviation calculation formula for the generated physical spatial domain two-dimensional discrete Gaussian kernel function is as follows:

[0040]

[0041] in, The standard deviation of the two-dimensional discrete Gaussian kernel function in the physical space domain. The physical calibration constant representing the imaging sensor is determined based on the attenuation coefficient of the sensor's modulation transfer function at the Nyquist frequency. Represents the physical scale relative scaling factor. Representative take and The larger of the two values; based on the generated Gaussian kernel function, perform two-dimensional discrete spatial convolution smoothing on the data matrix within the real physical reading window, and call the bilinear interpolation function after eliminating Nyquist aliasing artifacts; the Nyquist aliasing artifacts are false textures generated during the resampling process due to the spatial frequency exceeding the sampling limit.

[0042] As a preferred technical solution of the present invention, after using the bounding box topological inclusion relationship to perform cross-validation to filter out anomalies, cross-process geometric deduplication is performed on the set of valid feature coordinates that have passed the bounding box cross-validation to eliminate physical space coordinate collision redundancy caused by overlapping edges of layered grid tiles; the cross-process geometric deduplication is the processing logic for eliminating duplicate feature coordinates generated by overlapping grid tiles under multi-threaded parallel processing; the physical space coordinate collision redundancy is redundant data formed by multiple approximate feature coordinates corresponding to the same ground feature; the mathematical formula for performing feature cleaning logic by combining physical space Euclidean distance and feature coordinate confidence response value is as follows:

[0043]

[0044] in, This represents the set of high-precision feature points after deduplication. This represents the set of valid feature coordinates that have passed bounding box cross-validation. This is a set separator used to limit the selection criteria for elements in a set. The "belongs to" symbol indicates that an element belongs to the corresponding set. and These represent any two feature coordinates within the set of valid feature coordinates. This represents the absolute physical space Euclidean distance between two feature coordinates, obtained by inverse mapping in the absolute pixel coordinate system of the entire image. This represents the preset absolute physical space Euclidean distance rejection threshold. Represents the logical OR operator. and These represent the high-dimensional local response confidence scores of the deep learning matching model's forward inference output, which are attached to the feature coordinates. The symbol represents the greater than or equal to relation.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] This invention utilizes a hierarchical recursive grid generation module to perform multi-level logical segmentation and normalized scaling feature extraction of images. It then employs coordinate inverse kinematics and random sampling consensus algorithms to aggregate matching point pairs across levels, obtaining initial homography transformation parameters. This approach eliminates significant macroscopic initial geometric distortions between cross-source remote sensing images with low computational redundancy, effectively establishing a global coarse registration benchmark. After preliminary correction based on the initial homography transformation parameters, a relative scaling factor is calculated according to the physical resolution of the multi-source high-resolution reference image and the physical resolution of the image to be registered. Combined with a preset set of field-of-view magnifications, a dynamic field-of-view feature pyramid is constructed in the underlying data pipeline. This forces a unified resampling alignment of the image data matrix, placing the physical scale alignment pressure at the data input of the deep learning matching model. This overcomes the scale variance trap introduced by differences in physical ground sampling distances between heterogeneous sensors, improving the matching accuracy of heterogeneous image features.

[0047] In the distributed concurrent inference phase, a multi-threaded distributed concurrent engine loads multi-scale slice tasks, uses a model shell interceptor to capture local temporary index information, and performs addressing reconstruction using a full-batch global identifier matrix to replace the local temporary index information. Then, combined with bounding box topological inclusion relationship cross-validation to filter out abnormal points, a dense feature point set with unique absolute global coordinates across processes is generated, which solves the problems of tensor local index aliasing and global geographic coordinate system inaccuracy in large-scale parallel processing.

[0048] This invention employs an adaptive anti-aliasing discrete low-pass filtering mechanism that dynamically generates a Gaussian kernel function based on a physical scale relative scaling factor. This kernel performs convolutional smoothing on the resampling window to eliminate Nyquist aliasing artifacts, ensuring the purity of feature extraction. A cross-process high-precision feature deduplication mechanism combines physical spatial Euclidean distance and feature coordinate confidence response values ​​for feature cleaning, eliminating coordinate collision redundancy caused by overlapping edges of layered grid slices. This improves the robustness of the global spatial transformation matrix solution and the final alignment accuracy. The overall scheme achieves end-to-end optimization from macro-level coarse-grained to micro-level fine-grained alignment, significantly improving the processing efficiency and geographic topological confidence of large-scale remote sensing image registration. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0050] Figure 1 is a flowchart of the method of the present invention.

[0051] Figure 2 is a flowchart of the hierarchical recursive mesh coarse registration process of the present invention.

[0052] Figure 3 is a flowchart of the physical scale alignment and dynamic field-of-view feature pyramid construction sub-process of the present invention.

[0053] Figure 4 is a flowchart of the global identifier addressing and dense feature verification sub-process of the present invention. Detailed Implementation

[0054] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0055] The following is in conjunction with the appendix Figures 1-4 The embodiments of the present invention will be described in detail below.

[0056] Example 1: This example discloses a remote sensing image registration method based on hierarchical recursive coarse matching and physical scale alignment. This example processes multi-source high-resolution remote sensing image data, specifically performing the following steps:

[0057] In step one, the input multi-source high-resolution reference image and the image to be registered are analyzed using an affine transformation model to read the geographic metadata and affine transformation parameters of the multi-source high-resolution reference image and the image to be registered. The six-parameter matrix of geographic metadata is represented as follows: Arbitrary pixel space coordinates Projected onto real geophysical space coordinates The mathematical formula is:

[0058]

[0059] in, Represents the absolute geographic spatial x-coordinate, Represents the absolute geographic spatial vertical coordinate. and The absolute geographic coordinates representing the origin at the top left corner of the remote sensing image. and The horizontal and vertical ground sampling distances of remotely sensed images, i.e., pixel spatial resolution. and Characterizing the rotational skew of remotely sensed images, Represents the x-coordinate of pixels. Represents the vertical coordinate of a pixel.

[0060] In step two, a hierarchical recursive mesh generation module is used to construct a multi-level mesh for logical segmentation and normalized scaling feature extraction of the multi-source high-resolution reference image and the image to be registered. The target input image size is set to... Construct a hierarchical grid set For the first The layers are meshed, and the multi-source high-resolution reference image and the image to be registered are cut into layers. The formula for calculating the physical pixel size of each sub-image is as follows:

[0061]

[0062] in, Represents the pixel value of the sub-image height. Represents the width of the sub-image in pixels. Represents the height dimension of the target input image. Represents the width of the target input image. Representing the Number of horizontal grid divisions per layer Representing the Number of vertical grid divisions per layer This represents the floor function. Next, the cropped sub-image... Forced to use bilinear interpolation function Mapping to the standard normalized target size of the deep learning matching model The mathematical formula is:

[0063]

[0064] in, Represents the normalized target size image. This represents the cropped sub-image. Represents the pixel value of the sub-image height. Represents the width of the sub-image in pixels. Represents the number of image channels. Represents a bilinear interpolation function. Represents the standard normalized target size, typically The value is either 512 pixels or 1024 pixels. The normalized target-size image is input into a deep learning feature extraction backbone network (a neural network used to extract feature points and descriptors from the normalized image) to extract a set of matching point pairs. ,in Represents the index of the matching point within the sub-image. This represents the total number of matching points extracted from a single sub-image. It utilizes the local grid offsets of each layer of the grid. The normalized coordinates, along with their corresponding scaling factors, are mapped back to the full-image absolute pixel coordinate system of the multi-source high-resolution reference image or the image to be registered. The mathematical formula is as follows:

[0065]

[0066] in, and Represents the absolute pixel coordinates of the restored full image. and The normalized x and y coordinates are extracted by the deep learning matching model. Represents the pixel value of the sub-image height. Represents the width of the sub-image in pixels. Represents the standard normalized target size. and Represents the horizontal and vertical offset of the local mesh. This represents the Hadamard product operator, i.e., element-wise multiplication; the scaling factor is determined by... and Definition. Next, aggregate all absolute pixel coordinate pairs of the entire image generated by all hierarchical grid levels, and use the random sampling consensus algorithm to solve for the global homography matrix. To remove outlier coordinates, the mathematical formula is:

[0067]

[0068] in, Represents the homogeneous representation of matched coordinates on multi-source high-resolution reference images. This represents the homogeneous representation of the matching coordinates on the image to be registered. Represents the global homography matrix. This represents the individual matrix elements in the global homography matrix. The initial homography transformation parameters, i.e., the global homography matrix, are obtained through cross-level aggregation calculations using coordinate inverse kinematics and a random sampling consensus algorithm to eliminate large initial field-of-view offsets. .

[0069] In step three, using the multi-source high-resolution reference image as a reference, the initial homography transformation parameters, i.e., the global homography matrix, are utilized. Preliminary geometric correction is performed on the image to be registered to initially align it with the multi-source high-resolution reference image in pixel space. Then, a relative scaling factor is calculated based on the physical resolutions of the multi-source high-resolution reference image and the image to be registered. Combined with a preset set of field-of-view magnifications, a dynamic field-of-view feature pyramid is constructed in the underlying data pipeline. Finally, a forced uniform resampling alignment is performed on the preliminarily corrected multi-source high-resolution reference image data matrix and the image to be registered data matrix. The physical resolution of the multi-source high-resolution reference image is then read. Physical resolution of the image to be registered Calculate the physical scale relative scaling factor The mathematical formula is:

[0070]

[0071] in, Represents the physical scale relative scaling factor. The physical resolution representing a multi-source high-resolution reference image. This represents the physical resolution of the image to be registered. This represents the function that takes the maximum value. This represents the function that takes the minimum value. If the reference image's physical resolution is... Smaller than the physical resolution of the image to be registered Then, the scaling factor of the multi-source high-resolution reference image is assigned. Scaling factor of the image to be registered Set the standard input tensor size of the model to... The formula for calculating the actual physical read window size dynamically extracted from the front-end data stream is:

[0072]

[0073] in, The actual physical readout window size represents the multi-source high-resolution reference image. Represents the actual physical readout window size of the image to be registered. The size of the standard input tensor of the representative model. The scaling factor representing the multi-source high-resolution reference image. The scaling factor represents the image to be registered. This represents a floor operation. Next, it is combined with a preset set of field-of-view magnifications. ,in Represents the set of field of view magnifications. Represents the number of field magnifications. Representing the Each field of view magnification. The physical width of the actual physical read window is forcibly magnified in the underlying data pipeline. To construct a dynamic field-of-view feature pyramid, a bilinear interpolation function is used. Remodeling A normalized tensor of the size covers the input model; wherein, each field of view magnification in the preset set of field of view magnifications is no greater than the physical scale relative scaling factor. .

[0074] In step four, for the images that have undergone preliminary correction and unified resampling alignment, a multi-threaded distributed concurrent engine is scheduled to load multi-scale tiling tasks. While the multi-threaded distributed concurrent engine is loading multi-scale tiling tasks, each valid tile pushed into the queue is assigned a unique global grid identifier across processes, constructing a full-batch global identifier matrix. ,in Represents batch size. This represents a set of integers. A model shell interceptor captures local temporary index information from the feature queue output by the forward inference of the deep learning matching model. Addressing reconstruction is performed using the full-batch global identifier matrix to replace the local temporary index information. Cross-validation using bounding box topological inclusion relationships filters outliers, generating a dense set of feature points with unique absolute global coordinates across processes. When data is fed into the deep learning matching model for forward inference, a tensor queue with local temporary index information output by the deep learning matching model is intercepted. ,in This represents the total number of feature points extracted in the current batch. A one-dimensional dimensionality reduction mapping is performed on the global identifier matrix of the entire batch mounted in GPU memory to reconstruct it into absolute geographic identity labels. The mathematical formula is:

[0075]

[0076] in, Representing the The absolute geographic identity label corresponding to each feature point Represents the global identifier matrix for the entire batch. Representing the Each feature point contains local temporary index values. This represents the total number of feature points extracted in the current batch. It represents a logical universal quantifier, referring to any value within a range. This refers to the feature point index. Next, a batch consistency verification function is constructed for each extracted feature coordinate. By comparing the absolute pixel coordinates of the entire image after inverse kinematics with the absolute geographic identity tags Cross-validation of the topological inclusion relationships of the bounding boxes is performed, and the real-time outlier detection and filtering formula is as follows:

[0077]

[0078] in, Representing the The validity determination result of each feature coordinate. Representing the Each feature coordinate, Represents absolute geographical identity tags The associated real geographic grid bounding box boundary function.

[0079] In step five, the effective feature point set generated in step four and filtered by bounding box cross-validation is used to solve the global spatial transformation matrix using the least squares method. Based on the global spatial transformation matrix, global pixel resampling is performed on the image to be registered, and a high-resolution aligned remote sensing image is output.

[0080] Example 2: This example further elaborates on the data flow under the distributed concurrent architecture based on Example 1. In the face of high-concurrency cutting of multi-scale slice tasks loaded by the multi-threaded distributed concurrent engine, the system starts the multi-threaded asynchronous loading engine to schedule the dynamic multi-scale slices.

[0081] Before feeding the computational stream tensor into the deep learning matching model for concatenation, this embodiment pre-instantiates a contiguous address space in the host memory pool to mount the full batch global identifier matrix. The model shell interceptor uses a lock-free asynchronous queue to capture in real time the sequence of local temporary index information released by the feature queue of the deep learning matching model's forward inference output, and synchronously flattens the high-dimensional calculation results in the GPU memory. It uses the local temporary index of the tensor as a hash key to retrieve the corresponding absolute geographic identity tag in the memory pool, enabling the heterogeneous tensor coordinate systems between concurrent processes to be accurately projected and restored to the original absolute geographic coordinate system with extremely low computational power consumption. This fundamentally eliminates the risk of batch data coordinate crosstalk and local contamination in parallel processing.

[0082] Example 3: This example compares and verifies the specific performance differences in image registration between the remote sensing image registration method based on hierarchical recursive coarse matching and physical scale alignment and the existing implicit learning scheme that does not include physical ground sampling distance benchmark alignment.

[0083] Set up a detailed comparison:

[0084] The comparative model employs a feature pyramid implicit learning network with random fixed step size pruning at the input. Dynamic forced resampling based on physical ground sampling distance differences is not performed within the data loading pipeline. The model feature tensors are directly extracted and globally matched within the pixel space.

[0085] The experiment selected pairs of high-frequency repetitive textured farmland remote sensing images covering the same geographical area but originating from different aircraft sensors (with physical resolutions differing by a factor of 5).

[0086] The method in Example 1 first performs dynamic front-end field-of-view magnification compensation on the data (the set of field-of-view magnifications is...). It forces uniform resampling alignment in the underlying data pipeline; while the contrast ratio directly feeds heterogeneous slices into the deep learning matching model.

[0087] Table 1 shows a comparison of the experimental parameters.

[0088] Comparative Example none none 14.57 pixels 1204 3450 milliseconds Example 1 have have 1.82 pixels 8956 1210 milliseconds

[0089] As shown in Table 1, compared with the comparative example, the remote sensing image registration method based on hierarchical recursive coarse matching and physical scale alignment uses resampling based on physical ground sampling distance alignment and an anomaly point anti-aliasing interception detection function to eliminate high-frequency mismatch points caused by sensor physical differences. While acquiring large-scale dense feature point pairs, it reduces the registration geometric error to an extremely low level, which greatly improves the processing efficiency and geographic topological confidence of fully automated high-precision aligned images.

[0090] Example 4: Based on the underlying architecture of Example 2, this example constructs an adaptive anti-aliasing filtering mechanism for physical space domain feature purification and a cross-process high-precision feature deduplication mechanism.

[0091] In this embodiment, the optical remote sensing satellite sensor captures photon energy in orbit and quantizes it into a discrete digital matrix. The ground sampling distance reflected by the digital matrix follows the Shannon-Nyquist spatial sampling limit theory. In the forced uniform resampling alignment stage of the underlying data pipeline, real-world physical readout windows with different physical coverage are mapped to the model's standard input tensor size. At that time, the physical scale of the target data matrix is ​​relative to the scaling factor. At a high level, it indicates that the source image is in a compressed downsampled channel.

[0092] The periodicity of high-frequency spatial signals exceeds the carrying capacity limit of the resampled target grid. Complex surface textures exceeding the Nyquist frequency are mirrored and folded on the interpolated grid, forming moiré artifacts and false jagged edges unrelated to real ground features. The shallow convolutional receptive field of the deep learning matching model extracts these high-frequency energies as illusory feature responses. An adaptive anti-aliasing discrete low-pass filtering mechanism transforms this problem into a physical process with a closed-loop analytical solution. The physical scale is scaled relative to the target grid through mathematical mapping. The control parameter for the low-pass cutoff frequency in the physical space domain is transformed into a parameter. The system dynamically instantiates an adaptive low-pass filter based on the current absolute geographic identity tag of the segment. The standard deviation of the two-dimensional discrete Gaussian kernel function in the physical space domain is generated. Constrained by mathematical definitions, the formula is:

[0093]

[0094] in, The standard deviation of the two-dimensional discrete Gaussian kernel function in the physical space domain. The physical calibration constant representing the imaging sensor is determined based on the attenuation coefficient of the sensor's modulation transfer function at the Nyquist frequency. Represents the physical scale relative scaling factor. Representative take and The larger of the two values. When the physical scale is relative to the scaling factor. When the value exceeds a preset threshold, the adaptive anti-aliasing filter is triggered; the preset threshold can be set to 1.2, which indicates that when the physical scale relative scaling factor is greater than 1.2, there is significant downsampling.

[0095] Physical calibration constant The attenuation rate at the Nyquist frequency is set based on the optical modulation transfer function curve of the multi-source high-resolution reference image acquisition satellite at the time of its manufacture. The standard deviation is read by the system's underlying computing unit. Then, it requests a vector cache block from the host memory allocator, generating a block with a size specification of [size not specified]. A two-dimensional discrete weight kernel, where This represents the rounding up operation. The streaming multiprocessor built into the graphics processing unit activates the video memory bandwidth, transforming the two-dimensional discrete weight kernel into a sliding matrix and applying it to the data matrix surface within the real physical readout window. Pixels at the image physical boundaries are extended outward using an alternating odd-even mirror fill technique. Low-frequency skeleton data streams are filtered through adaptive anti-aliasing smoothing and then processed using a bilinear interpolation function. Standardized tensor arrays eliminate the scale variance hidden by physical scale transformations. A cross-process high-precision feature deduplication mechanism is used to resolve spatial geometric overlap issues caused by topology partitioning in multi-threaded distributed concurrent engines.

[0096] When the hierarchical recursive mesh generation module performs logical segmentation, it introduces overlapping bounding boxes of predetermined widths for connected local meshes. These overlapping bounding boxes ensure semantic integrity at the seams of the segments. The same real-world feature corner point is segmented into multiple independent tensor batches and fed in parallel to the deep learning matching model for processing. After the model shell interceptor replaces local temporary index information with absolute geographic identity labels, duplicate feature corner points appear as redundant clusters with approximate coordinate values ​​in the absolute pixel coordinate system of the entire map. Inputting these redundant clusters into the homography parameter estimator of the random sampling consensus algorithm causes the calculated geometric model to only fit the distortion state around the mesh seams. Feature cleaning logic based on the memory penetration characteristics of the full-batch global identifier matrix blocks this evolution.

[0097] After outliers are filtered out by bounding box topological cross-validation, the set of valid feature coordinates is as follows. Residing in a shared cache pool of a non-uniform memory access architecture. A multi-threaded concurrency scheduler initiates a multi-level pointer scan thread pool to traverse all feature pair combinations. The absolute physical space Euclidean distance between two feature coordinates. Monitored by the computing unit. If the absolute physical space Euclidean distance is less than or equal to the preset absolute physical space Euclidean distance removal threshold. The arbitration gateway is activated. The arbitration gateway reads high-dimensional local response confidence scores attached to feature coordinates across different levels. and The mathematical formula for the cleaning logic is:

[0098]

[0099] in, This represents the set of high-precision feature points after deduplication. This represents the set of valid feature coordinates that have passed bounding box cross-validation. and These represent any two feature coordinates within the set of valid feature coordinates. This represents the absolute physical space Euclidean distance between two feature coordinates, obtained by inverse mapping in the absolute pixel coordinate system of the entire image. This represents the preset absolute physical space Euclidean distance rejection threshold. Represents the logical OR operator. and These represent the confidence scores of the high-dimensional local response, respectively. This represents a greater than or equal to relation. Coordinate nodes with higher signal response strength are granted pointer retention permissions. The high-precision feature point set exhibits isotropic physical spatial distribution characteristics, increasing the probability of successfully capturing globally correct interior points through consistent random sampling iterations.

[0100] Stress simulation verification was performed using the Earth Observation System Data Center. The test environment was equipped with multiple Tensor Core graphics accelerator cards and interconnect architecture. The test input source consisted of multi-source high-resolution reference images, with a physical resolution calibration of 0.8 meters and a data dimension of [missing information]. Absolute pixels. The physical resolution of the image to be registered is calibrated to 5.8 meters, and the data dimension is... Absolute pixels. Physical scale relative scaling factor. The value is 7.25. The system generates multi-scale local grid tensors for the task. Satellite physical calibration constants. The value is set to 0.48, which is determined based on the attenuation coefficient of the satellite sensor modulation transfer function measured at the Nyquist frequency.

[0101] Standard deviation of the two-dimensional discrete Gaussian kernel function in the physical space domain The calculated output is 3.45. The parallel stream processor performs convolutional smoothing and noise filtering on tensor blocks. In the cross-process high-precision feature deduplication stage, the concurrent engine outputs 2.3 million local temporary feature coordinates. The memory scheduler controller calls the full batch global identifier matrix to complete the tensor inverse reprojection within 42 milliseconds. A preset absolute physical space Euclidean distance removal threshold is set. The baseline length is 2.5 times the absolute pixel distance. The Euclidean distance scan tree removes 457,000 overlapping collision coordinates caused by concurrent cutting redundancy.

[0102] The random sampling consensus algorithm module accepts a high-precision set of feature points, and the iteration time is stable within 100 milliseconds. The ablation experiment evaluation parameters and index results are shown in Table 2. Anti-aliasing discrete low-pass filtering and cross-process geometric deduplication were not enabled in the baseline group.

[0103] Ablation group A independently activates the anti-aliasing discrete low-pass filter mechanism.

[0104] Ablation group B activates two deep architectural mechanisms.

[0105] The measurement scale uses the point cloud of standard geographic control points extracted by the photogrammetric workstation as the reference data.

[0106] Table 2:

[0107] Reference group operating status 28.6% 457,118 2150 milliseconds 4.28 absolute pixels Ablation Group A Operating Status 3.7% 485,322 2890 milliseconds 1.65 absolute pixels Ablation Group B Operating Status 3.2% 0 95 milliseconds 0.41 absolute pixels

[0108] Table 2 shows that, due to the frequency bottleneck of downsampling, 28.6% of the deep learning features in the baseline group were high-frequency forged coordinates, and the root mean square geometric error of absolute orthophoto registration was 4.28 absolute pixels. After ablation group A activated physical space smoothing control, the proportion of high-frequency forged feature points extracted decreased to 3.7%. Due to the increased sensitivity of the model in processing clean textures, ablation group A captured more redundant overlapping coordinates at mesh boundaries, and the convergence time of global homography matrix calculation increased to 2890 milliseconds. After ablation group B activated the feature cleaning logic, the number of redundant collision and overlapping coordinates was reduced to zero. Confidence arbitration based on signal response intensity preserved high-confidence geometric features, and the convergence time of the global transformation solution network was reduced to 95 milliseconds, with a registration error of 0.41 absolute pixels. The mechanism itself ensures the technical implementation of large-scale resolution image alignment.

[0109] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0110] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A remote sensing image registration method based on hierarchical recursive coarse-to-fine registration and physical scale alignment, characterized in that, Includes the following steps: Step 1: Read the geographic metadata and Earth affine transformation parameters of the multi-source high-resolution reference image and the image to be registered; Step 2: Using the hierarchical recursive mesh generation module, a multi-level mesh is constructed to logically segment and extract normalized scaling features from the multi-source high-resolution reference image and the image to be registered. The initial homography transformation parameters for eliminating the large initial field of view offset are obtained by cross-level aggregation calculation using coordinate inverse solution and random sampling consistency algorithm. Step 3: Using the multi-source high-resolution reference image as a reference, after performing preliminary correction on the image to be registered according to the initial homography transformation parameters, calculate the relative scaling factor based on the physical resolution of the multi-source high-resolution reference image and the physical resolution of the image to be registered. Combined with the preset field of view magnification set, construct a dynamic field of view feature pyramid in the underlying data pipeline, and perform forced unified resampling alignment on the multi-source high-resolution reference image data matrix and the image to be registered data matrix. Step 4: For the images that have completed preliminary correction and unified resampling alignment, schedule a multi-threaded distributed concurrent engine to load multi-scale slicing tasks. Capture local temporary index information of the feature queue output by the forward inference of the deep learning matching model through the model shell interceptor. Use the full batch global identifier matrix to perform addressing reconstruction to replace the local temporary index information. Use the bounding box topological inclusion relationship to perform cross-validation to filter out anomalies and generate a dense feature point set with unique absolute global coordinates across processes. The global spatial transformation matrix is ​​calculated based on a dense feature point set. Global pixel resampling is then performed on the image to be registered to output a high-resolution aligned remote sensing image.

2. The remote sensing image registration method based on hierarchical recursive coarse matching and physical scale alignment according to claim 1, characterized in that, In step one, the input multi-source high-resolution reference image and the image to be registered are analyzed by the earth affine transformation model, and the geographic metadata six-parameter matrix is represented as , the mathematical formula for projecting any pixel space coordinates to real earth physical space coordinates is: wherein, represents an absolute geographical space horizontal coordinate, represents an absolute geographical space vertical coordinate, and characterizes the absolute geographical coordinates of the upper left corner origin of the remote sensing image, and characterizes the horizontal and vertical ground sampling distance of the remote sensing image, i.e. the pixel space resolution, and characterizes the rotation angle of the remote sensing image, represents a pixel horizontal coordinate, represents a pixel vertical coordinate.

3. The hierarchical recursive coarse-to-fine and physical scale alignment based remote sensing image registration method according to claim 1, wherein, In step two, the height and width dimensions of the target input image are set, and a hierarchical set consisting of multi-level layered grids is constructed. Each level of layered grid has a corresponding number of horizontal grid segments and a number of vertical grid segments. For any level of grid division, the multi-source high-resolution reference image and the image to be registered are cut into sub-images multiplied by the number of horizontal sub-segments and the number of vertical sub-segments of the current level grid. The height pixel value of each sub-image is the floor value of the target input image height divided by the number of horizontal sub-segments of the current level grid, and the width pixel value of each sub-image is the floor value of the target input image width divided by the number of vertical sub-segments of the current level grid.

4. The remote sensing image registration method based on hierarchical recursive coarse matching and physical scale alignment according to claim 3, characterized in that, In step two, the cropped sub-image is scaled to the standard normalized target size required by the deep learning matching model using a bilinear interpolation function, forming a normalized target size image.

5. The remote sensing image registration method based on hierarchical recursive coarse matching and physical scale alignment according to claim 4, characterized in that, In step two, the normalized target size image is input into the deep learning feature extraction backbone network to extract a set of matching point pairs. Using the local grid horizontal offset and local grid vertical offset of each layer grid, as well as the corresponding scaling factor, the normalized horizontal and vertical coordinates are mapped back to the absolute pixel horizontal and vertical coordinates of the entire image of the multi-source high-resolution reference image or the image to be registered. The scaling factor is the ratio of the sub-image width pixel value to the standard normalized target size and the ratio of the sub-image height pixel value to the standard normalized target size.

6. The remote sensing image registration method based on hierarchical recursive coarse matching and physical scale alignment according to claim 5, characterized in that, In step two, all absolute pixel coordinate point pairs generated by the layered grid are aggregated, and the global homography matrix is ​​solved using the random sampling consensus algorithm. The global homography matrix satisfies the mapping relationship between the homogeneous expression of matching coordinates on the multi-source high-resolution reference image and the homogeneous expression of matching coordinates on the image to be registered. Abnormal coordinate points are removed by the random sampling consensus algorithm to obtain the initial homography transformation parameters.

7. The remote sensing image registration method based on hierarchical recursive coarse matching and physical scale alignment according to claim 1, characterized in that, In step three, the physical resolution of the reference image of the multi-source high-resolution reference image and the physical resolution of the image to be registered are read, and the physical scale relative scaling factor is calculated. The value of the physical scale relative scaling factor is the larger of the physical resolution of the reference image and the physical resolution of the image to be registered, divided by the smaller value.

8. The remote sensing image registration method based on hierarchical recursive coarse matching and physical scale alignment according to claim 7, characterized in that, In step three, if the physical resolution of the reference image is smaller than the physical resolution of the image to be registered, then the scaling factor of the multi-source high-resolution reference image is assigned equal to the physical scale relative scaling factor, and the scaling factor of the image to be registered is equal to 1.0; the standard input tensor size of the model is set. The formula for calculating the actual physical read window size dynamically extracted from the front-end data stream is: in, The actual physical readout window size represents the multi-source high-resolution reference image. Represents the actual physical readout window size of the image to be registered. The size of the standard input tensor of the representative model. The scaling factor representing the multi-source high-resolution reference image. The scaling factor represents the image to be registered. This represents the floor operation.

9. The remote sensing image registration method based on hierarchical recursive coarse matching and physical scale alignment according to claim 8, characterized in that, In step three, based on a preset set of field-of-view magnifications, the physical extent of the real physical reading window is forcibly magnified by a corresponding factor in the underlying data pipeline to construct a dynamic field-of-view feature pyramid. This pyramid is then reshaped into a standardized tensor of standard size using a bilinear interpolation function to cover the input model. Each field-of-view magnification in the preset set of field-of-view magnifications is no greater than the physical scale relative scaling factor.

10. The remote sensing image registration method based on hierarchical recursive coarse matching and physical scale alignment according to claim 1, characterized in that, In step four, when the multi-threaded distributed concurrent engine loads the multi-scale slicing task, a unique cross-process global grid identifier is assigned to each valid slice pushed into the queue, and a full-batch global identifier matrix is ​​constructed. ,in Represents batch size. Represents a set of integers.

Images