An optical remote sensing image uniform dense homonymy point matching method, system, device and storage medium based on dense optical flow

Abstract

The application discloses a dense optical flow-based optical remote sensing image uniform dense matching method and system, equipment and a storage medium, and belongs to the technical field of satellite remote sensing image processing. The application solves the technical problem that the prior art lacks a dense optical flow-based optical remote sensing image uniform dense matching method which can recognize sufficient and uniformly distributed matching points in various complex scenes and has high positioning accuracy. The application confirms a remote sensing image overlap area, divides the remote sensing image overlap area into a plurality of remote sensing image blocks, constructs a displacement field solving model and a Gaussian image pyramid, obtains a displacement field of the remote sensing image overlap area, acquires a pre-matching matching point pair set, distributes the pre-matching matching point pair set into corresponding remote sensing image overlap area image blocks, and screens a final matching point set. The application is used for realizing a dense optical flow-based optical remote sensing image uniform dense matching method which can recognize sufficient and uniformly distributed matching points in various complex scenes and has high positioning accuracy.

Description

Technical Field

[0001] This invention relates to the field of satellite remote sensing image processing technology, specifically to a method, system, device, and storage medium for matching uniformly dense corresponding points in optical remote sensing images based on dense optical flow. Background Technology

[0002] Common methods for matching corresponding points in high spatial resolution optical remote sensing satellite imagery can be broadly categorized into deep learning methods and traditional algorithms based on a "feature extraction, description, matching, and mismatch removal" strategy. Data-driven deep learning methods are highly dependent on training samples, and existing unsupervised learning neural networks suffer from poor accuracy. Supervised learning neural networks, on the other hand, require large amounts of high-precision manually labeled datasets. Furthermore, for the problem of labeling corresponding points in optical remote sensing imagery spanning large areas, long time spans, and under different imaging conditions, they suffer from problems such as a massive amount of labeling, high recognition difficulty, and overly limited scene coverage. Traditional algorithms driven by the "extraction, description, and matching" strategy perform well in regions with prominent features and stable structural information. However, they often struggle to identify corresponding points in regions with weak texture, complex local distortions, and large differences in grayscale appearance. Moreover, constrained by multi-stage matching strategies, even if sufficient identical features can be extracted from two images, if the feature description algorithm cannot accurately construct a discriminative feature space or if the pre-defined consistency model deviates significantly from the actual spatial mapping, traditional algorithms cannot provide a sufficient number of corresponding points. In addition, existing traditional algorithms are mostly designed for prominent geometric features in images. For complex land cover types such as mountainous and urban areas, their recognition results often show a high degree of spatial clustering.

[0003] In the prior art, Chinese patent document CN103279935A discloses a "Method and System for Super-Resolution Reconstruction of Thermal Infrared Remote Sensing Images Based on MAP Algorithm," which acquires a sequence of thermal infrared remote sensing images, including at least two frames. It employs a high-precision automatic registration method based on corner features, utilizing automatic extraction and matching of corner points to complete registration. The MAP algorithm is used to achieve super-resolution reconstruction of the image sequence, and an adaptive selection method for the edge penalty function threshold is proposed for the selection of potential function parameters in the Gibbs model. The reconstructed target resolution image is then evaluated for application-oriented quality. However, this technical solution, which relies on automatic extraction and matching of corner points for registration, often struggles to identify corresponding points in areas with weak texture, complex local distortions, and large differences in grayscale appearance. Furthermore, due to the limitations of multi-stage matching strategies, the identification results for complex terrain types such as mixed mountainous and urban areas often exhibit a high degree of spatial clustering.

[0004] In summary, the existing technology lacks a technical solution for a method of matching uniformly dense corresponding points in optical remote sensing images based on dense optical flow that can identify a sufficient number of evenly distributed corresponding points in various complex scenarios with high positioning accuracy. Summary of the Invention

[0005] This invention solves the technical problem of the lack of a method for matching uniformly dense corresponding points in optical remote sensing images based on dense optical flow that can identify a sufficient number of uniformly distributed corresponding points in various complex scenarios and has high positioning accuracy.

[0006] The present invention discloses a method for matching uniformly dense corresponding points in optical remote sensing images based on dense optical flow, comprising the following steps:

[0007] Step 1: Based on the geographic coordinate information of multiple remote sensing images, identify the overlapping area of ​​remote sensing images, divide the overlapping area of ​​remote sensing images into multiple remote sensing image blocks, and obtain the image blocks of the overlapping area of ​​remote sensing images.

[0008] Step 2: Construct the displacement field solution model and Gaussian image pyramid respectively to obtain the displacement field of the overlapping area of ​​the remote sensing images;

[0009] Step 3: Design a sparse same-name point sampling strategy to obtain a pre-matched same-name point pair set;

[0010] Step 4: Assign the corresponding points in the pre-matched set of corresponding points to the corresponding image blocks in the overlapping area of ​​the remote sensing images, filter out the correctly matched corresponding points in the overlapping area of ​​the remote sensing images, and take the set of correctly matched corresponding points as the final set of matched corresponding points.

[0011] Furthermore, in this embodiment of the invention, the size of the remote sensing image block in step 1 is no greater than 800 pixels * 800 pixels.

[0012] Furthermore, in this embodiment of the invention, step 2, which involves constructing a displacement field solution model and a Gaussian image pyramid to obtain the displacement field of image blocks in the overlapping area of ​​remote sensing images, specifically includes:

[0013] The size of the image patch in the overlapping area of ​​the remote sensing image is expanded outward. Based on the spatial grayscale information and gradient structure information of the image patch, a displacement field smoothing constraint is introduced to construct a displacement field calculation model, thereby obtaining the primary displacement field of the image patch in the overlapping area. After resampling and sampling parameter correction at the top layer of the Gaussian image pyramid, the primary displacement field of the image patch in the overlapping area is used as the initial value for calculating the displacement field of the next layer in the Gaussian image pyramid. The calculation result of the bottom layer in the Gaussian image pyramid is used as the expanded displacement field of the image patch in the overlapping area. The expanded part of the image patch in the overlapping area is removed, and the middle part of the image patch in the overlapping area is used as the displacement field of the image patch in the overlapping area.

[0014] Furthermore, in this embodiment of the invention, the sparse homonym sampling strategy in step 3, which obtains a pre-matched homonym pair set, specifically includes:

[0015] Extract a set of candidate corner points from a remote sensing image. Based on the minimum distance between adjacent points, divide the overlapping area of ​​the remote sensing image into multiple adjacent grids. Select the corner point closest to the center of each adjacent grid from the set of candidate corner points. Based on the displacement field of the overlapping area of ​​the remote sensing image, obtain the corresponding points of the corner point closest to the center of the adjacent grid, and obtain a set of pre-matched corresponding point pairs.

[0016] Furthermore, in this embodiment of the invention, the extraction of a set of candidate corner points from a remote sensing image specifically involves:

[0017] Corner points are extracted from remote sensing images. The displacement field of the overlapping area of ​​remote sensing images is uniformly sampled to obtain a supplementary point set of remote sensing images. The corner points of remote sensing images and the supplementary point set of remote sensing images are combined to obtain a candidate corner point set of remote sensing images.

[0018] The present invention discloses a system for matching uniformly dense corresponding points in optical remote sensing images based on dense optical flow, comprising the following modules:

[0019] The segmentation module identifies overlapping areas of remote sensing images based on geographic coordinate information from multiple remote sensing images, and divides these overlapping areas into multiple remote sensing image blocks to obtain the image blocks of the overlapping areas.

[0020] The module is used to construct the displacement field solution model and the Gaussian image pyramid respectively, so as to obtain the displacement field of the overlapping area of ​​the remote sensing image.

[0021] The matching module designs a sparse sampling strategy for corresponding points to obtain a set of pre-matched corresponding point pairs;

[0022] The filtering module assigns the corresponding points in the pre-matched set of corresponding points to the corresponding image blocks in the overlapping area of ​​the remote sensing images, filters out the correctly matched corresponding points in the overlapping area of ​​the remote sensing images, and uses the set of correctly matched corresponding points as the final set of matched corresponding points.

[0023] The electronic device of the present invention includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus.

[0024] Memory, used to store computer programs;

[0025] When the processor executes the program stored in the memory, it implements any of the above-described methods for matching uniformly dense corresponding points in optical remote sensing images based on dense optical flow.

[0026] The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-described methods for matching uniformly dense corresponding points in optical remote sensing images based on dense optical flow.

[0027] This invention solves the technical problem of existing technologies lacking a method for matching uniformly dense corresponding points in optical remote sensing images based on dense optical flow, which can identify a sufficient number of uniformly distributed corresponding points in various complex scenarios with high positioning accuracy. Specific beneficial effects include:

[0028] This invention addresses the problem of corresponding point identification in high spatial resolution optical remote sensing images by proposing a uniform and dense corresponding point matching method based on dense optical flow. The method coarsely locates overlapping areas of remote sensing images using geographic coordinate information from multiple images. By combining spatial grayscale and gradient structure information from the images, a displacement field calculation model is constructed to calculate the spatial correspondence between two remote sensing images pixel by pixel. A Gaussian image pyramid is used to address the large displacement problem between the images. Corner point extraction is employed to detect the corner structure in the images, providing a more accurate description of the spatial location and geometric structure features. Grid sampling is used to uniformly sample corner points, improving the uniformity of the corner point set. The corresponding corresponding points are then determined using the displacement field of the overlapping areas. A random sampling consensus algorithm based on a local affine model is used to eliminate mismatched corresponding point pairs. This method addresses the potential for severe local geometric inconsistencies between remote sensing images from different time periods and the overfitting phenomenon in the displacement field, thereby achieving high-precision uniform and dense matching of corresponding points in complex grayscale and geometrically distorted scenes. This method for matching uniformly dense corresponding points in optical remote sensing images based on dense optical flow can extract sufficient, uniformly distributed, and densely packed high-precision corresponding points in various complex image scenarios. Attached Figure Description

[0029] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0030] Figure 1 This is a schematic diagram of the image homonym point extraction results of the urban construction area described in Implementation Method 1;

[0031] Figure 2 This is a schematic diagram of the results of the extraction of corresponding points in mountainous images as described in Implementation Method 1;

[0032] Figure 3 This is a schematic diagram of the results of extracting corresponding points in the plain area as described in Implementation Method 1;

[0033] Figure 4 This is a schematic diagram of the results of extracting corresponding points in the mixed terrain area as described in Implementation Method 1. Detailed Implementation

[0034] Various embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. The embodiments described with reference to the drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0035] Implementation Method 1. The method for matching uniformly dense corresponding points in optical remote sensing images based on dense optical flow described in this implementation method includes the following steps:

[0036] Step 1: Based on the geographic coordinate information of multiple remote sensing images, identify the overlapping area of ​​remote sensing images, divide the overlapping area of ​​remote sensing images into multiple remote sensing image blocks, and obtain the image blocks of the overlapping area of ​​remote sensing images.

[0037] Step 2: Construct the displacement field solution model and Gaussian image pyramid respectively to obtain the displacement field of the overlapping area of ​​the remote sensing images;

[0038] Step 3: Design a sparse same-name point sampling strategy to obtain a pre-matched same-name point pair set;

[0039] Step 4: Assign the corresponding points in the pre-matched set of corresponding points to the corresponding image blocks in the overlapping area of ​​the remote sensing images, filter out the correctly matched corresponding points in the overlapping area of ​​the remote sensing images, and take the set of correctly matched corresponding points as the final set of matched corresponding points.

[0040] In this embodiment, the size of the remote sensing image block in step 1 is no greater than 800 pixels * 800 pixels.

[0041] In this embodiment, step 2, which involves constructing a displacement field calculation model and a Gaussian image pyramid to obtain the displacement field of image blocks in the overlapping area of ​​remote sensing images, specifically includes:

[0042] The size of the image patch in the overlapping area of ​​the remote sensing image is expanded outward. Based on the spatial grayscale information and gradient structure information of the image patch, a displacement field smoothing constraint is introduced to construct a displacement field calculation model, thereby obtaining the primary displacement field of the image patch in the overlapping area. After resampling and sampling parameter correction at the top layer of the Gaussian image pyramid, the primary displacement field of the image patch in the overlapping area is used as the initial value for calculating the displacement field of the next layer in the Gaussian image pyramid. The calculation result of the bottom layer in the Gaussian image pyramid is used as the expanded displacement field of the image patch in the overlapping area. The expanded part of the image patch in the overlapping area is removed, and the middle part of the image patch in the overlapping area is used as the displacement field of the image patch in the overlapping area.

[0043] In this embodiment, the sparse homonym sampling strategy in step 3, which obtains the pre-matched homonym pair set, specifically includes:

[0044] Extract a set of candidate corner points from a remote sensing image. Based on the minimum distance between adjacent points, divide the overlapping area of ​​the remote sensing image into multiple adjacent grids. Select the corner point closest to the center of each adjacent grid from the set of candidate corner points. Based on the displacement field of the overlapping area of ​​the remote sensing image, obtain the corresponding points of the corner point closest to the center of the adjacent grid, and obtain a set of pre-matched corresponding point pairs.

[0045] In this embodiment, the extraction of a set of candidate corner points from a remote sensing image specifically involves:

[0046] Corner points are extracted from remote sensing images. The displacement field of the overlapping area of ​​remote sensing images is uniformly sampled to obtain a supplementary point set of remote sensing images. The corner points of remote sensing images and the supplementary point set of remote sensing images are combined to obtain a candidate corner point set of remote sensing images.

[0047] Existing technologies lack a technical problem: a method for matching uniformly dense corresponding points in optical remote sensing images based on dense optical flow that can identify a sufficient number of evenly distributed corresponding points in various complex scenarios with high positioning accuracy.

[0048] To address the aforementioned technical problems, this embodiment provides a method for matching uniformly dense corresponding points in optical remote sensing images based on dense optical flow. This method solves the problem of identifying corresponding points between optical remote sensing images of the same spatial resolution at different time phases. In this embodiment, remote sensing images at different time phases are represented by Img1 and Img2, respectively. The specific steps include:

[0049] Step 1: Based on the geographic coordinate information of multiple remote sensing images, identify the overlapping area of ​​remote sensing images, divide the overlapping area of ​​remote sensing images into multiple remote sensing image blocks, and obtain the image blocks of the overlapping area of ​​remote sensing images.

[0050] Step 1 involves coarse localization of overlapping areas in remote sensing images based on geographic coordinate information and a strategy for dividing overlapping areas. To reduce computational load and improve the efficiency of the same-point identification method, it is first necessary to determine the approximate overlapping range of the remote sensing images based on the geographic coordinate information accompanying the image release. Using the geographic coordinates of the starting point at the top left corner of the remote sensing image and other geographic coordinate information such as spatial resolution, a mapping relationship between remote sensing image pixels and geographic coordinates can be established, i.e.:

[0051] (1)

[0052] (2)

[0053] in, The geographic coordinates of the starting point at the top left corner of the image. For image spatial resolution parameters, For pixels Image plane coordinates, For pixels The geographical coordinates.

[0054] The boundary positions of the remote sensing image in the four directions of east, west, south, and north can be calculated using formula (1), that is, the four-sided range of the remote sensing image. Wherein, the four-sided range of Img1 is... The four-dimensional range of Img2 is At this point, Img1 and Img2 are unified under the same geographic coordinate system, thus allowing the calculation of the overlapping area of ​​the remote sensing images. If there is no overlapping area in the remote sensing images, the identification process ends directly, and the corresponding points are identified as an empty set. If there is an overlapping area in the remote sensing images, the subsequent steps continue.

[0055] Since uncontrolled sensor positioning struggles to achieve high-precision positioning, relying solely on geographic coordinates from remote sensing images is insufficient to accurately determine the location of overlapping areas. To minimize the impact of non-overlapping areas on subsequent methods, the overlapping areas need to be reduced by a certain size, resulting in the following range: The overlapping area of ​​the remote sensing image can be mapped to the grid coordinate system of Img1 by formula (2) to obtain the overlapping area of ​​Img1.

[0056] If the overlapping area is too large, it will severely affect the computational efficiency and accuracy of dense optical flow. Therefore, the Img1 overlapping area is uniformly divided into multiple remote sensing image blocks, each no larger than 800 pixels * 800 pixels. When calculating the displacement field, the image blocks are used as the computational unit.

[0057] Step 2: Construct the displacement field solution model and Gaussian image pyramid respectively to obtain the displacement field of the overlapping area of ​​the remote sensing images;

[0058] The size of the image patch in the overlapping area of ​​the remote sensing image is expanded outward. Based on the spatial grayscale information and gradient structure information of the image patch, a displacement field smoothing constraint is introduced to construct a displacement field calculation model, thereby obtaining the primary displacement field of the image patch in the overlapping area. After resampling and sampling parameter correction at the top layer of the Gaussian image pyramid, the primary displacement field of the image patch in the overlapping area is used as the initial value for calculating the displacement field of the next layer in the Gaussian image pyramid. The calculation result of the bottom layer in the Gaussian image pyramid is used as the expanded displacement field of the image patch in the overlapping area. The expanded part of the image patch in the overlapping area is removed, and the middle part of the image patch in the overlapping area is used as the displacement field of the image patch in the overlapping area.

[0059] Step 2 involves constructing a displacement field solution model based on dense optical flow. Influenced by imaging conditions and seasonal factors, remote sensing images from different time phases tend to exhibit significant nonlinear differences in spatial grayscale information. Constructing a displacement field solution model solely based on the assumption of constant illumination is insufficient to guarantee the stability of a method for matching uniformly dense corresponding points in optical remote sensing images using dense optical flow. Therefore, the method used in this implementation combines the spatial grayscale information and gradient structure information of the remote sensing image, introduces displacement field smoothing constraints, and jointly constructs a displacement field solution model. The expression for the displacement field solution model is as follows:

[0060] (3)

[0061] (4)

[0062] (5)

[0063] (6)

[0064] (7)

[0065] in, The gradient structure weights are set a priori. To smooth the constraint parameters, To avoid infinite decimals with a penalty term of 0, The value is typically taken as 0.001. = For the displacement field to be determined, For any pixel location in the image For gradient operators, For spatial grayscale information, For gradient structure information, The displacement field is subject to smoothing constraints. The energy function is minimized by solving the problem. The primary displacement field of the image block in the overlapping area of ​​the remote sensing image is obtained.

[0066] However, the primary displacement field of the overlapping image blocks in the remote sensing image calculated by the displacement field solution model has a large displacement problem, leading to inaccurate matching of corresponding points. To address the potential large displacement problem between Img1 and Img2, a multi-level solution is performed using a pyramid image strategy, specifically:

[0067] First, the downsampling parameters of the remote sensing images are predefined manually. With maximum allowable displacement The default values ​​are 0.75 and 30, respectively. Then, the number of pyramid levels is calculated according to the following formula. And construct a Gaussian image pyramid:

[0068] (8)

[0069] Assuming the size of the lowest layer remote sensing image is For the first Layered pyramid image, its size is The displacement field is calculated level by level from the top of the pyramid. The calculation result of the current layer is used as the initial value for the displacement field calculation of the next layer after resampling and sampling parameter correction. In addition, considering that the accuracy of the displacement field usually decreases significantly in the edge region, it is necessary to expand each image block by a certain size as the bound before solving it. In this embodiment, the bound is set to 30 pixels by default. After the calculation is completed, the expanded part is discarded, and only the middle part is retained as the displacement field of the current image block. The calculation result of the bottom layer image is the displacement field of the image block in the overlapping area of ​​the remote sensing images. After calculating for all image blocks, the displacement field of the overlapping area of ​​the remote sensing images Img1 and Img2 is obtained.

[0070] However, using the displacement field of the overlapping area of ​​remote sensing images to match corresponding points in remote sensing images will lead to problems of image radiation and geometric distortion. Therefore, a sparse corresponding point sampling strategy is proposed to solve the problems of image radiation and geometric distortion.

[0071] Step 3: Design a sparse same-name point sampling strategy to obtain a pre-matched same-name point pair set;

[0072] Extract a set of candidate corner points from a remote sensing image. Based on the minimum distance between adjacent points, divide the overlapping area of ​​the remote sensing image into multiple adjacent grids. Select the corner point closest to the center of each adjacent grid from the set of candidate corner points. Based on the displacement field of the overlapping area of ​​the remote sensing image, obtain the corresponding points of the corner point closest to the center of the adjacent grid, and obtain a set of pre-matched corresponding point pairs.

[0073] Step 3: Design a sparse corresponding point sampling strategy based on the corner point structure of remote sensing images. Based on the grayscale information and gradient structure information of the overlapping area of ​​remote sensing images, establish the spatial correspondence of corresponding points in two remote sensing images and perform sparse corresponding point sampling.

[0074] Corner structural features typically exhibit good saliency and stability. On one hand, corner structural features are easily assessed using posterior methods such as manual judgment to determine whether matching corresponding points is a correct match. On the other hand, corner structures generally show good stability against image radiometry and geometric distortion, such as road intersections and building vertices. Compared to random point selection, corners can more accurately describe the spatial location and geometric structure features of remote sensing images.

[0075] This implementation first extracts Img1 corner points block by block using the Harris operator to obtain the corner point set. Considering that corner structures are prone to being undetectable or excessively clustered in mountainous and urban areas, a supplementary corner set is obtained by uniformly sampling the displacement field of the overlapping area of ​​the remote sensing images. To maximize the extraction of corner points, supplementary points should be added in areas with weak textures or other areas where corner points are difficult to extract. The sampling radius should be... This is a priori setting value, with a default value of 30. and Together they form the set of candidate corner points .

[0076] To improve the uniformity of the point set, the set of points to be selected needs further filtering. Define the minimum distance between adjacent points. Divide the overlapping areas of remote sensing images into multiple evenly distributed regions as much as possible. For grids of varying sizes, those with insufficient dimensions at the edges can be merged with adjacent grids or left untreated. If a set of candidate corner points exists within a grid... From multiple corner points in the image, retain the corner point closest to the center of the grid and delete the other corner points. The coordinates of the retained corner points can be obtained from the displacement field of the overlapping area of ​​the remote sensing image in step 2.

[0077] In Img1, any corner point The image coordinates are Displacement field At this corner point The pixel value at that location is represented as the displacement. Therefore, Img2 contains... Corresponding points The image coordinates are .

[0078] By using a sparse homonym sampling strategy to filter out candidate points that are too close together, a pre-matched homonym pair set can be obtained. .in, For the first For points with the same name, and These are its coordinates in the Img1 and Img2 pixel coordinate systems, respectively.

[0079] Step 4: Assign the corresponding points in the pre-matched set of corresponding points to the corresponding image blocks in the overlapping area of ​​the remote sensing image, filter out the correctly matched corresponding points in the overlapping area of ​​the remote sensing image, and take the set of correctly matched corresponding points as the final set of matched corresponding points.

[0080] Step 4 involves a method for removing mismatched corresponding points based on a local affine model. Due to objective factors such as imaging angle and changes in ground elevation, remote sensing images from different time phases may exhibit severe local geometric inconsistencies. For example, farmland, overpasses, and buildings in the same area may show varying degrees and directions of geometric deformation. Furthermore, the displacement field calculated in Step 2 for the overlapping area of ​​the remote sensing images is prone to overfitting, i.e., erroneous stretching. Additionally, shadows and fog can also affect the accuracy of displacement field calculations in the overlapping area. Therefore, further filtering of the pre-matched corresponding point pair set in Step 3 is necessary to remove mismatched point pairs.

[0081] This implementation uses a random sampling consensus algorithm to pre-match the set of identical point pairs. The iteration count and model deviation parameters can be manually adjusted, with default values ​​of 1000 and 3.0 respectively. Considering the large size of remote sensing images and the significant differences in local geometric distortion in different regions, the global model estimated using partial points cannot accurately represent the mapping relationship between points and cannot determine the correctness of the matching of corresponding points.

[0082] Therefore, this embodiment will Each point is assigned to an image block in the overlapping area of ​​the remote sensing images divided in step 1. Each image block in the overlapping area of ​​the remote sensing images is independently filtered using RANSAC. Finally, the filtering results of each image block in the overlapping area of ​​the remote sensing images are integrated into the same set to obtain the final set of matching points with the same name.

[0083] This embodiment proposes a method for matching uniformly dense corresponding points in optical remote sensing images based on dense optical flow. By combining the spatial grayscale information and gradient structure information of the remote sensing images and introducing smoothness constraints, the spatial correspondence of overlapping areas of remote sensing images at different time phases is established pixel by pixel to obtain the displacement field of the overlapping area of ​​the remote sensing images. Then, based on the displacement field, the sparse corresponding point sampling strategy and local affine model proposed in this embodiment are used for sampling and screening to finally obtain a high-precision matching corresponding point set.

[0084] Compared to commonly used "feature extraction, description, and matching" methods, this implementation method can identify a sufficient number of evenly distributed corresponding points in various complex scenarios with high positioning accuracy. Compared to deep learning algorithms, this implementation method does not require a large amount of manpower and material resources, and has good universality and robustness to complex nonlinear grayscale mapping and local geometric distortions in images.

[0085] To better illustrate the method for matching uniformly dense corresponding points in optical remote sensing images based on dense optical flow described in this embodiment, the following examples provide a detailed description:

[0086] Example 1. In this example, based on the spatial grayscale information and gradient structure information of the joint remote sensing image, the spatial correspondence of the overlapping area of ​​the remote sensing image is established pixel by pixel to obtain the displacement field of the overlapping area of ​​the remote sensing image. Then, the corresponding points and mismatched points are further screened and filtered out through the sparse corresponding point sampling strategy, and finally, a sufficient number of corresponding points with uniform distribution are obtained.

[0087] Four sets of satellite remote sensing images acquired by the "Jilin-1" high-resolution satellite were used as the experimental targets. Each set of remote sensing images underwent preprocessing steps such as orthorectification, panchromatic multispectral fusion, and cropping, and were all standard TIF format raster images. Based on the overlapping observation areas, the four sets of satellite remote sensing images were divided into: "Mountainous Area Group," "Urban Area Group," "Plains Group," and "Complex Terrain Group." The "Complex Terrain Group" refers to complex mixed terrain areas including mountains, urban areas, and water bodies. Based on the different coverage areas, they can be divided into mountainous remote sensing images, urban area remote sensing images, plains remote sensing images, and mixed terrain remote sensing images. The resolution, size, number of bands, and pixel grayscale value of the remote sensing images did not affect the results of this embodiment. For ease of processing and description, the experimental data were uniformly processed: the spatial resolution was 0.75 meters, the size was uniformly cropped to 5000*5000, the number of bands was 4, and the pixel grayscale value was... .

[0088] This embodiment utilizes satellite remote sensing imagery acquired by the "Jilin-1" high-resolution satellite for testing, achieving large-scale, automated, sufficient, and evenly distributed identification of corresponding points. For remote sensing imagery in different scenarios, this embodiment can accurately match a sufficient number of evenly distributed corresponding points, as shown in the specific results. Figure 1-4 As shown. This embodiment uses default parameters for the experiment.

[0089] Implementation Method 2. The optical remote sensing image uniform dense corresponding point matching system based on dense optical flow described in this implementation method includes the following modules:

[0090] The segmentation module identifies overlapping areas of remote sensing images based on geographic coordinate information from multiple remote sensing images, and divides these overlapping areas into multiple remote sensing image blocks to obtain the image blocks of the overlapping areas.

[0091] The module is used to construct the displacement field solution model and the Gaussian image pyramid respectively, so as to obtain the displacement field of the overlapping area of ​​the remote sensing image.

[0092] The matching module designs a sparse sampling strategy for corresponding points to obtain a set of pre-matched corresponding point pairs;

[0093] The filtering module assigns the corresponding points in the pre-matched set of corresponding points to the corresponding image blocks in the overlapping area of ​​the remote sensing images, filters out the correctly matched corresponding points in the overlapping area of ​​the remote sensing images, and uses the set of correctly matched corresponding points as the final set of matched corresponding points.

[0094] Implementation Method 3. An electronic device according to this implementation method includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0095] Memory, used to store computer programs;

[0096] When the processor executes the program stored in the memory, it implements the method for matching uniform and dense corresponding points in optical remote sensing images based on dense optical flow as described in Embodiment 1.

[0097] Implementation Method 4. The computer-readable storage medium described in this implementation method stores a computer program, which, when executed by a processor, implements the method for matching uniformly dense corresponding points in optical remote sensing images based on dense optical flow as described in Implementation Method 1.

[0098] The foregoing has provided a detailed description of the method, system, device, and storage medium for matching uniformly dense corresponding points in optical remote sensing images based on dense optical flow, as proposed in this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for matching uniform dense homonymy points of optical remote sensing images based on dense optical flow, characterized in that, Includes the following steps: Step 1: Based on the geographic coordinate information of multiple remote sensing images, identify the overlapping area of ​​remote sensing images, divide the overlapping area of ​​remote sensing images into multiple remote sensing image blocks, and obtain the image blocks of the overlapping area of ​​remote sensing images. Step 2: Construct the displacement field solution model and Gaussian image pyramid respectively to obtain the displacement field of the overlapping area of ​​the remote sensing images; Step 2, which involves constructing a displacement field solution model and a Gaussian image pyramid to obtain the displacement field of image blocks in the overlapping area of ​​remote sensing images, specifically involves: The size of the image block in the overlapping area of ​​the remote sensing image is expanded outward. Based on the spatial grayscale information and gradient structure information of the image block, displacement field smoothing constraints are introduced to construct a displacement field calculation model and obtain the primary displacement field of the image block in the overlapping area. After resampling and sampling parameter correction at the top layer of the Gaussian image pyramid, the primary displacement field of the image block in the overlapping area is used as the initial value for calculating the displacement field of the next layer in the Gaussian image pyramid. The calculation result of the bottom layer in the Gaussian image pyramid is used as the expanded displacement field of the image block in the overlapping area. The expanded part of the image block in the overlapping area is removed, and the middle part of the image block in the overlapping area is used as the displacement field of the image block in the overlapping area. Step 3: Design a sparse same-name point sampling strategy to obtain a pre-matched same-name point pair set; The sparse homonym sampling strategy in step 3, which obtains a pre-matched homonym pair set, specifically involves: Extract a set of candidate corner points from a remote sensing image. Based on the minimum distance between adjacent points, divide the overlapping area of ​​the remote sensing image into multiple adjacent grids. Select the corner point closest to the center of each adjacent grid from the set of candidate corner points. Based on the displacement field of the overlapping area of ​​the remote sensing image, obtain the corresponding points of the corner point closest to the center of the adjacent grid, and obtain a set of pre-matched corresponding point pairs. The extraction of a candidate corner point set from a remote sensing image specifically involves: The corner points of the remote sensing image are extracted, and the displacement field of the overlapping area of ​​the remote sensing image is uniformly sampled to obtain the supplementary point set of the remote sensing image. The corner points of the remote sensing image and the supplementary point set of the remote sensing image are combined to obtain the candidate corner point set of the remote sensing image. Step 4: Assign the corresponding points in the pre-matched set of corresponding points to the corresponding image blocks in the overlapping area of ​​the remote sensing images, filter out the correctly matched corresponding points in the overlapping area of ​​the remote sensing images, and take the set of correctly matched corresponding points as the final set of matched corresponding points.

2. The method according to claim 1, wherein, The size of the remote sensing image block in step 1 is no greater than 800 pixels * 800 pixels.

3. A dense optical flow based optical remote sensing image uniform dense homonymy point matching system, characterized in that, Includes the following modules: The segmentation module identifies overlapping areas of remote sensing images based on geographic coordinate information from multiple remote sensing images, and divides these overlapping areas into multiple remote sensing image blocks to obtain the image blocks of the overlapping areas. The module is used to construct the displacement field solution model and the Gaussian image pyramid respectively, so as to obtain the displacement field of the overlapping area of ​​the remote sensing image. The construction module includes building a displacement field solution model and a Gaussian image pyramid to obtain the displacement field of image blocks in the overlapping area of ​​remote sensing images, specifically: The size of the image block in the overlapping area of ​​the remote sensing image is expanded outward. Based on the spatial grayscale information and gradient structure information of the image block, displacement field smoothing constraints are introduced to construct a displacement field calculation model and obtain the primary displacement field of the image block in the overlapping area. After resampling and sampling parameter correction at the top layer of the Gaussian image pyramid, the primary displacement field of the image block in the overlapping area is used as the initial value for calculating the displacement field of the next layer in the Gaussian image pyramid. The calculation result of the bottom layer in the Gaussian image pyramid is used as the expanded displacement field of the image block in the overlapping area. The expanded part of the image block in the overlapping area is removed, and the middle part of the image block in the overlapping area is used as the displacement field of the image block in the overlapping area. The matching module designs a sparse sampling strategy for corresponding points to obtain a set of pre-matched corresponding point pairs; The sparse homonym sampling strategy in the matching module obtains a pre-matched homonym pair set, specifically as follows: Extract a set of candidate corner points from a remote sensing image. Based on the minimum distance between adjacent points, divide the overlapping area of ​​the remote sensing image into multiple adjacent grids. Select the corner point closest to the center of each adjacent grid from the set of candidate corner points. Based on the displacement field of the overlapping area of ​​the remote sensing image, obtain the corresponding points of the corner point closest to the center of the adjacent grid, and obtain a set of pre-matched corresponding point pairs. The extraction of a candidate corner point set from a remote sensing image specifically involves: The corner points of the remote sensing image are extracted, and the displacement field of the overlapping area of ​​the remote sensing image is uniformly sampled to obtain the supplementary point set of the remote sensing image. The corner points of the remote sensing image and the supplementary point set of the remote sensing image are combined to obtain the candidate corner point set of the remote sensing image. The filtering module assigns the corresponding points in the pre-matched set of corresponding points to the corresponding image blocks in the overlapping area of ​​the remote sensing images, filters out the correctly matched corresponding points in the overlapping area of ​​the remote sensing images, and uses the set of correctly matched corresponding points as the final set of matched corresponding points.

4. An electronic device, comprising: It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; The processor, when executing a program stored in memory, implements the method for matching uniformly dense corresponding points in optical remote sensing images based on dense optical flow as described in any one of claims 1-2.

5. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the method for matching uniformly dense corresponding points in optical remote sensing images based on dense optical flow as described in any one of claims 1-2.

Images