Graph theory-based seamless remote sensing image splicing method and system

Abstract

The application discloses a seamless remote sensing image splicing method and system based on graph theory, and relates to the technical field of image processing, and comprises the following steps: acquiring a set of to-be-processed multi-source remote sensing images, and preprocessing each to-be-processed multi-source remote sensing image in the set; determining intersecting n-gons between adjacent preprocessed multi-source remote sensing images, and obtaining a plurality of intersecting n-gons; determining the central axes of each intersecting n-gon, and generating bisectors between adjacent preprocessed multi-source remote sensing images based on the central axes; cutting each preprocessed multi-source remote sensing image based on the bisectors, and obtaining a target splicing line; and cutting and splicing each preprocessed multi-source remote sensing image, and obtaining a target seamless splicing image. Through implementation of the application, image segments that have undergone geometric correction and color unification are integrated into a complete target seamless splicing image according to geographical coordinates, and the problems of low data reuse rate and difficult cross-scene application are solved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a graph theory-based method and system for seamless stitching of remote sensing images. Background Technology

[0002] With the rapid development of high-resolution Earth observation technology, satellite remote sensing imagery has become a core data resource for refined monitoring in many fields such as power, land resources, and environmental protection. Particularly in industry applications such as power infrastructure management, there is an urgent need to build a one-stop service system encompassing data aggregation, storage management, and rapid processing to support critical business scenarios such as dense power transmission corridor inspections, flood season disaster prevention, and geological disaster early warning. Satellite remote sensing data, with its advantages of wide-area coverage, rapid response, and dynamic updates, plays an irreplaceable role in the industry's digital transformation process. For example, it can be used to identify the risk of external damage by regularly acquiring images of power transmission corridors, or to compare and analyze the construction progress of new energy power plants using multi-temporal imagery.

[0003] However, the raw imagery distributed by satellite data providers generally suffers from significant geometrical positional deviations and color inconsistencies, failing to directly meet the needs of specialized business applications. Geometrical deviations lead to decreased positioning accuracy for critical features such as power poles and lines, while color differences cause a noticeable "patchwork effect" after stitching together multiple images, severely impacting the accuracy of subsequent image interpretation and analysis.

[0004] Currently, the standardized processing of remote sensing imagery mainly relies on two approaches: outsourcing to third-party professional organizations and using commercial image processing software. While third-party services can meet general needs, they generally suffer from long data processing cycles and numerous workflow steps, making them unsuitable for applications with extremely high timeliness requirements, such as emergency monitoring and post-disaster assessment. Existing mainstream software still faces several technical bottlenecks when processing multi-source, large-scale remote sensing imagery: for example, stitching line generation is often based on simple geometric segmentation principles, making it difficult to effectively avoid man-made structures, leading to visual misalignment at stitching boundaries; color consistency processing often uses global histogram matching or single template correction, which can easily introduce local color shifts when fusing multi-source images, especially noticeable in areas with typical spectral characteristics such as vegetation and water bodies. Furthermore, the automation capabilities of such software are limited, still requiring significant manual intervention for large areas, making it difficult to adapt to the large-scale production pace of "weekly updates of meter-level images and monthly updates of sub-meter-level images."

[0005] Meanwhile, users in industries such as power have stringent requirements for the geometric accuracy and color reproduction of image products. For example, monitoring of power transmission channels requires that the geometric positioning error of images be controlled within the meter level, and flood monitoring needs to clearly distinguish the boundaries between water bodies and wetlands. However, the existing technology system lacks standardized processing indicators for industry applications, making it difficult to use image products from different sources and at different times in a collaborative manner, objectively creating "data silos."

[0006] In recent years, with the continuous deployment of domestic and international satellite constellations, the ability to acquire multi-source remote sensing data has been significantly enhanced. The data volume of a single scene image has reached the GB level, and image stitching over a large area often requires processing hundreds or thousands of scenes. Against this backdrop, traditional technologies face three main challenges: First, significant heterogeneity exists among multi-source images. Differences in the imaging mechanisms of different sensors lead to complex geometric distortion patterns, making it difficult for traditional correction methods to achieve uniform adaptation. Second, accuracy and efficiency are difficult to balance during stitching line generation. Existing graph-based partitioning algorithms are prone to path deviations in complex scenes such as densely built-up areas, still requiring manual correction. Third, the global optimization effect of color consistency is poor. The phenomenon of "different spectra for the same object" is common due to differences in imaging time, illumination, and atmospheric conditions, and existing color balancing techniques are unable to effectively suppress cumulative color differences over large areas. Furthermore, industry applications are increasingly demanding standardization in image processing, involving not only the unification of coordinate systems and file formats but also adaptation to the customized needs of different business scenarios. However, existing remote sensing image stitching technology still lacks a standardized processing framework covering the entire process, resulting in low reuse rate of data processing results and difficulties in cross-scenario applications. Summary of the Invention

[0007] In view of this, the present invention provides a graph theory-based method and system for seamless stitching of remote sensing images, in order to solve the problems of low data reuse rate and difficulty in cross-scenario application caused by the lack of a standardized processing framework for the entire process of related remote sensing image stitching technologies for power grid applications.

[0008] The present invention adopts the following technical solution.

[0009] A first aspect of the present invention provides a graph-theory-based method for seamless stitching of remote sensing images, the method comprising: A set of multi-source remote sensing images to be processed is obtained, and each multi-source remote sensing image in the set is preprocessed to obtain a set of preprocessed multi-source remote sensing images. The multi-source remote sensing images to be processed are orthophotos after geometric correction. The preprocessing includes performing down-biting processing first, and then performing color consistency processing. For the preprocessed multi-source remote sensing image set, the intersecting n-sided polygons between adjacent preprocessed multi-source remote sensing images are determined, resulting in several intersecting n-sided polygons, where n is a positive integer greater than or equal to 3. Determine the central axis of each of the intersecting n-gons, and generate the bisector between adjacent preprocessed multi-source remote sensing images based on the central axis; Based on the bisecting line, each of the preprocessed multi-source remote sensing images is cropped to obtain the target stitching line; Based on the target stitching line, each of the preprocessed multi-source remote sensing images is cropped, and the cropped images are stitched together to obtain a seamless target stitched image.

[0010] Optionally, the downsizing process includes: The multi-source remote sensing images to be processed are stretched using nonlinear Gamma stretching. Based on the normalized vegetation index and the normalized water index, the vegetation area and water area in each multi-source remote sensing image to be processed are enhanced respectively. The stretched and enhanced multi-source remote sensing images to be processed are converted into 8-bit multi-source remote sensing images to be processed.

[0011] Optionally, the color consistency processing is a hybrid radiometric correction, including: Candidate reference images are determined based on the geographic extent, acquisition time, and spatial resolution of the multi-source remote sensing images to be processed; Radiometric calibration and atmospheric correction are performed on the candidate reference images to obtain the surface reflectance products of the candidate reference images; The multi-source remote sensing image to be processed is resampled until the resolution of the multi-source remote sensing image to be processed and the surface reflectance product are consistent. The invariance probability of corresponding pixels between the multi-source remote sensing image to be processed and the surface reflectance product is calculated based on the IR-MAD algorithm. Based on the rich parameters of image information, the multi-source remote sensing image to be processed is divided into blocks, and each block is corrected based on the correction model constructed based on invariant probability. Based on the distance distribution from each pixel to its block center, a continuous correction parameter field is generated by weighted bilinear interpolation, and pixel-by-pixel color correction is performed on the multi-source remote sensing image to be processed according to the correction parameter field.

[0012] Optionally, the color consistency processing is a correction combining global and local aspects, including: The color correction parameters of each multi-source remote sensing image to be processed are uniformly calculated using a global optimization strategy. A local optimization strategy is used to eliminate residual color deviation in the overlapping areas of adjacent multi-source remote sensing images to be processed.

[0013] Optionally, the color consistency processing is a correction based on a color reference library, including: The grayscale values ​​of the multi-source remote sensing images to be processed are mapped to the grayscale mean and standard deviation range of the target reference image through linear transformation. The target reference image is extracted from a preset color reference library.

[0014] Optionally, determining the intersecting n-gon between adjacent preprocessed multi-source remote sensing images includes: Geometric correction is performed on the preprocessed multi-source remote sensing images to determine their effective range; The vertex coordinates of the effective range are parsed from the metadata of the preprocessed multi-source remote sensing image to form a polygon describing the effective range, and the boundary coordinates of the effective range are obtained. Based on the effective range boundary coordinates, a spatial indexing algorithm is used to retrieve adjacent preprocessed multi-source remote sensing images whose effective ranges overlap with those of the preprocessed multi-source remote sensing images. Polygon Boolean intersection operation is used to calculate the effective range polygons of the preprocessed multi-source remote sensing image and adjacent preprocessed multi-source remote sensing images, and the closed intersecting n-sided polygon is obtained by solving. Traverse each preprocessed multi-source remote sensing image in the preprocessed multi-source remote sensing image set, and remove intersecting n-gons with an area smaller than a preset threshold to obtain several intersecting n-gons.

[0015] Optionally, generating the bisector between adjacent preprocessed multi-source remote sensing images based on the central axis includes: Identify the intersection points of the effective range boundaries of adjacent preprocessed multi-source remote sensing images on intersecting n-sided polygons, and select the two farthest intersection points as the start and end points respectively; Arrange the vertices of the intersecting n-sided polygons in order to form a sequence of vertices, and take the broken line segment between the starting point and the ending point as the central axis; Based on image segmentation, identify artificial building regions within intersecting n-sided polygons and generate corresponding binary masks; When the central axis passes through a building area, the central axis will be offset away from the building by a preset distance; Curve fitting and path simplification are performed on the offset central axis to generate the bisector.

[0016] Optionally, the setting of the preset distance includes: Based on the spatial resolution of the preprocessed multi-source remote sensing images, a baseline offset is determined. The reference offset is dynamically adjusted based on the pixel size of the artificial building area in the binary mask. The larger the pixel size of the building area, the greater the adjustment range of the reference offset. The adjusted offset is converted into an actual distance value, and the offset direction is determined by combining it with the geometric orientation of the central axis to complete the setting of the preset distance.

[0017] Optionally, obtaining the target splicing line includes: Use the effective range of the current preprocessed multi-source remote sensing image as the initial cropping region; Integrate the bisecting lines between the current preprocessed multi-source remote sensing images and adjacent preprocessed multi-source remote sensing images to construct a bisecting line network; The preprocessed multi-source remote sensing images are cropped sequentially using a bisecting line network, preserving the centroid region; Traverse each preprocessed multi-source remote sensing image in the preprocessed multi-source remote sensing image set to generate several Voronoi polygons; Check several Voronoi polygons to determine if there are any conflicts between them. If there are conflicts, adjust the bisecting network and re-trim until there are no conflicts between the several Voronoi polygons. Define the common boundary of several Voronoi polygons as the target splicing line.

[0018] A second aspect of the present invention provides a graph-based seamless remote sensing image stitching system for performing the graph-based seamless remote sensing image stitching method described in the first aspect of the present invention, comprising: The module includes a preprocessing module, a determining module, a generating module, a cropping module, and a stitching module, among which: The preprocessing module is used to acquire a set of multi-source remote sensing images to be processed, and to preprocess each multi-source remote sensing image in the set of multi-source remote sensing images to be processed to obtain a preprocessed set of multi-source remote sensing images. The preprocessing includes downsampling and color consistency processing. The multi-source remote sensing images to be processed are orthophotos after geometric correction. The determining module is used to determine the intersecting n-sided polygons between adjacent preprocessed multi-source remote sensing images for the preprocessed multi-source remote sensing image set, and obtain several intersecting n-sided polygons, where n is a positive integer greater than or equal to 3; The generation module is used to determine the central axis of each of the intersecting n-gons, and generate the bisector between adjacent preprocessed multi-source remote sensing images based on the central axis. The cropping module is used to crop each of the preprocessed multi-source remote sensing images based on the bisector to obtain the target stitching line; The stitching module is used to crop each of the preprocessed multi-source remote sensing images based on the target stitching line, and to stitch the cropped images together to obtain a target seamless stitched image.

[0019] Compared with the prior art, the beneficial effects of the present invention include at least the following: 1. The graph theory-based seamless remote sensing image stitching method provided by this invention acquires geometrically corrected orthophotos as a set of multi-source remote sensing images to be processed, ensuring that the images have accurate geographic coordinate references, eliminating geometric offsets of the original images, and performing down-biting and color consistency processing on the images. This solves the problem that high-bit images are difficult to display on PCs or web platforms, and eliminates color deviations caused by different satellite sensors and shooting conditions. It effectively avoids stitching errors caused by inconsistent data formats or color differences, while meeting the special requirements for image geometric accuracy and color consistency in power grid applications.

[0020] 2. The graph theory-based seamless remote sensing image stitching method provided by this invention accurately identifies the boundaries of overlapping areas between adjacent multi-source remote sensing images by determining the intersecting n-gons between them. This clarifies the overlapping parts that need to be focused on during the stitching process, providing accurate geometric range and boundary conditions for subsequent central axis calculation and bisector generation. It can effectively eliminate interference from non-overlapping areas, reduce invalid calculations, and improve processing efficiency. At the same time, it ensures that the subsequently generated stitching lines are completely located within the overlapping area, avoiding stitching misalignment or gaps caused by blurred boundaries. It is particularly suitable for processing complex overlapping relationships of large-scale multi-source images, providing accurate spatial positioning basis for achieving global seamless stitching.

[0021] 3. The graph theory-based seamless image stitching method provided by this invention determines the central axis of each intersecting n-gon and generates bisectors between adjacent images based on the central axis. Utilizing the property that the distance from the polygon's central axis to the boundary is equal, it ensures that the bisectors are always distributed along the central path of the overlapping area, mathematically guaranteeing the rationality of the stitching lines. Simultaneously, combined with image feature characteristics, the central axis automatically avoids key features such as high-rise buildings and towers during the generation process, preventing stitching misalignment caused by traditional stitching lines passing through redundant features. The generated bisectors possess both continuity and smoothness, providing a stable reference boundary for subsequent cropping and stitching, significantly improving the geometric accuracy and visual effect of the stitching results, and meeting the requirements for image detail integrity in power grid inspection.

[0022] 4. The graph theory-based seamless remote sensing image stitching method provided by this invention clips each image to be processed based on the bisecting lines between adjacent images, generating a target stitching line composed of Voronoi polygon boundaries. This achieves precise division of the stitching area for each multi-source remote sensing image to be processed, ensuring that all Voronoi polygons do not overlap and completely cover the entire area, avoiding the reuse or omission of overlapping areas, and allowing each multi-source remote sensing image to contribute only its best area to the stitching. At the same time, the target stitching line inherits the ground feature avoidance characteristics of the bisecting lines, further ensuring the rationality of the stitching boundary, providing a clear and executable basis for subsequent stitching operations, effectively improving the automation level and processing efficiency of large-scale image stitching, and adapting to the large-scale production needs of "weekly updates of meter-level images and monthly updates of sub-meter-level images" in power grid operations.

[0023] 5. The graph theory-based seamless image stitching method provided by this invention crops the preprocessed image based on the target stitching line and then stitches the cropped image together. This integrates geometrically corrected and color-unified image segments into a complete seamless image based on geographic coordinates, fully leveraging the advantages of the high-precision data source and optimized stitching line in the previous steps. The stitching result not only achieves seamless alignment in geometric position but also maintains natural consistency in color transition, meeting the needs of large-scale, high-precision imagery in scenarios such as UHV dense channel inspections, flood control and typhoon response, and geological disaster early warning in power grid applications. Furthermore, the seamless image can be further processed into sectional images according to power grid standards, providing unified and reliable spatial data support for subsequent line planning, hazard identification, and emergency decision-making, thus promoting the deep application of satellite remote sensing data in the digital transformation of the power grid. Attached Figure Description

[0024] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0025] Figure 1 This is a flowchart illustrating a graph-based method for seamless stitching of remote sensing images according to an embodiment of the present invention. Figure 2 This is a flowchart of hybrid radiation correction according to an embodiment of the present invention; Figure 3 This is a flowchart of global and local color consistency processing according to an embodiment of the present invention; Figure 4This is a flowchart of color consistency processing based on a color reference library according to an embodiment of the present invention; Figure 5 This is a structural block diagram of a graph-based seamless remote sensing image stitching system according to an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0027] According to Embodiment 1 of the present invention, a graph theory-based method for seamless stitching of remote sensing images is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0028] This embodiment provides a graph theory-based method for seamless stitching of remote sensing images. Figure 1 This is a flowchart of a graph-based seamless image stitching method for remote sensing images according to an embodiment of the present invention, as follows: Figure 1 As shown, the process includes the following steps: Step S101: Obtain a set of multi-source remote sensing images to be processed, and preprocess each multi-source remote sensing image in the set to obtain a preprocessed set of multi-source remote sensing images. The multi-source remote sensing images to be processed are orthophotos after geometric correction. The preprocessing includes performing down-biting processing first, and then performing color consistency processing.

[0029] Specifically, after acquiring the original images, the original band data that matches the spectral range of blue, green, red, and near-infrared bands is first selected based on the band response functions of various satellite sensors. For cases where some sensors lack corresponding bands, the missing blue, green, red, and near-infrared band data is generated through adjacent band interpolation or based on a spectral reconstruction model. This ensures that subsequent operations such as geometric correction and color consistency processing are all carried out based on the unified image data of the four blue, green, red, and near-infrared bands.

[0030] The multi-source remote sensing image set to be processed originates from raw images acquired by various types of satellite sensors. These raw images are captured and transmitted back by satellites and distributed by ground stations. Due to geometric positional deviations and color differences, they cannot be directly used for operational applications. They need to be converted into orthorectified images with accurate geographic coordinates through geometric correction steps such as feature point extraction, adaptive image matching, regional network adjustment, and orthorectification. The specific geometric correction process is as follows: First, unique and stable feature points are extracted from the images. Then, corresponding points are found between different images through adaptive image matching. Next, regional network adjustment technology is used to calculate these points to optimize the spatial location parameters of the images. Finally, orthorectification is used to eliminate the distortion caused by terrain undulation and image tilt, so that the pixels of each image can correspond to the real geographic coordinates.

[0031] The down-biting process in preprocessing converts 16-bit images into 8-bit images. After down-biting, color consistency processing is performed to eliminate color differences between different multi-source remote sensing images to be processed.

[0032] In some alternative embodiments, the downsizing process includes: Step a1: The nonlinear Gamma stretching method is used to stretch each of the multi-source remote sensing images to be processed.

[0033] Furthermore, nonlinear Gamma stretching is a processing method that nonlinearly maps image grayscale by adjusting the Gamma value. The Gamma value is dynamically set according to the image's brightness distribution. For areas with low brightness, use a Gamma value less than 1 to enhance details in dark areas; For areas with high brightness, a Gamma value greater than 1 is used to suppress grayscale overflow in overly bright areas. This avoids color saturation in some areas caused by linear stretching and makes the overall grayscale distribution of the image more in line with human visual habits, so that the histogram presents a normal distribution.

[0034] Step a2: Based on the normalized vegetation index and the normalized water index, the vegetation area and water area in each multi-source remote sensing image to be processed are enhanced.

[0035] Furthermore, the Normalized Difference Vegetation Index (NDVI) accurately identifies vegetated areas by calculating the difference in reflectance between the near-infrared and red light bands in the image; the Normalized Difference Water Index (NDIA) highlights water areas by analyzing the difference in reflectance between the green and near-infrared bands. During processing, vegetated and water areas are delineated based on the numerical ranges of these two indices. For vegetated areas, the proportion of the near-infrared band in the green light channel is increased to enhance the green hue; for water areas, the weights of blue and green light are adjusted to enhance the clarity of the water. This targeted enhancement makes vegetation brighter and water clearer, meeting the needs of power grid applications for identifying features around power lines.

[0036] Step a3: Convert the stretched and enhanced multi-source remote sensing image to be processed into an 8-bit multi-source remote sensing image.

[0037] Furthermore, the original remote sensing images are mostly 16-bit, which cannot be directly displayed on computers or web pages, and the large data volume is not conducive to efficient processing. When converting to 8-bit imagery, the stretching and enhancement results from the previous two steps are combined, and the adjusted grayscale values ​​are mapped proportionally to the 8-bit range to ensure that details in key areas such as vegetation and water bodies are not lost. The amount of data after conversion is significantly reduced, while maintaining the visual quality and ground feature identification of the image.

[0038] Specifically, if the multi-source remote sensing images to be processed have clear and high-precision surface reflectance products, radiometric correction methods based on surface reflectance products should be given priority. Using their accurate reflectance data as a color reference can ensure the physical authenticity of colors to the greatest extent. When there are high-resolution images of multiple time phases and the same area, and only the color differences between images caused by shooting conditions need to be eliminated, relative radiometric correction can be used to achieve color consistency through constant probability weight calculation and regression model construction. When it is necessary to accurately restore the radiation of ground objects from a physical level, and at the same time optimize color transitions by combining relative correction, hybrid radiometric correction is adopted. That is, first, absolute radiometric correction is performed using medium-resolution images to obtain basic radiometric information, and then relative radiometric correction is carried out using it as a reference, taking into account both physical accuracy and color consistency between images.

[0039] In some alternative embodiments, Figure 2 This is a flowchart of hybrid radiometric correction according to an embodiment of the present invention. Color consistency processing is hybrid radiometric correction, specifically including: Step b1: Determine candidate reference images based on the geographic extent, acquisition time, and spatial resolution of the multi-source remote sensing images to be processed.

[0040] Furthermore, in terms of geographical scope, the candidate reference images must cover the entire area of ​​the multi-source remote sensing images to be processed to ensure spatial consistency of the correction; In terms of timing, the shooting times for both should be as close as possible to minimize the impact of seasonal changes and differences in lighting conditions. In terms of spatial resolution, the resolution of the reference image needs to be compatible with the resolution of the multi-source remote sensing image to be processed (e.g., if the resolution of the image to be processed is 1 meter, the reference image can be selected with a medium resolution of 30 meters to facilitate subsequent scale matching).

[0041] By screening through these three dimensions, we can ensure that the candidate reference images can serve as reliable color benchmarks.

[0042] Step b2 involves radiometric calibration and FLAASH (Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes) atmospheric correction of the candidate reference image to obtain the surface reflectance product of the candidate reference image.

[0043] Furthermore, radiometric calibration converts the raw grayscale values ​​of the candidate reference image into radiance values ​​received by the sensor, eliminating the influence of differences in sensor response. FLAASH atmospheric correction uses specialized algorithms to eliminate atmospheric scattering and absorption interference, restoring the true reflectance characteristics of ground features. After these two steps, the candidate reference image is converted into a surface reflectance product—this product directly reflects the spectral characteristics of the ground features themselves, unaffected by external factors such as the atmosphere and sensors, becoming the "standard template" for color correction of the multi-source remote sensing image to be processed, ensuring that the corrected multi-source remote sensing image can accurately reflect the colors of ground features.

[0044] Step b3: Resample the multi-source remote sensing image to be processed until the resolution of the multi-source remote sensing image to be processed and the surface reflectance product are consistent. Use the IR-MAD (Iteratively Reweighted Multivariate Alteration Detection) algorithm to calculate the invariance probability of the corresponding pixels of the multi-source remote sensing image to be processed and the surface reflectance product respectively.

[0045] Furthermore, resampling involves adjusting the resolution of the multi-source remote sensing image to be processed using methods such as interpolation to ensure consistency with the surface reflectance product. For example, a 1-meter image to be processed might be resampled to 30 meters to ensure direct comparison at the pixel scale. The IR-MAD algorithm is used to identify "spectrally stable pixels" in the multi-source remote sensing image and the surface reflectance product. The spectral characteristics of these pixels are unaffected by time or environment, and their invariance probability value is closer to 1, indicating greater stability. This measure of pixel spectral consistency provides a basis for subsequent calibration model construction. Step b4: Based on the image information enrichment parameters, the multi-source remote sensing image to be processed is divided into blocks. Based on the correction model, each block of the multi-source remote sensing image to be processed is corrected. The correction model is constructed based on invariant probability. The image information enrichment parameters are obtained by calculating the information entropy of each block of the multi-source remote sensing image to be processed.

[0046] Furthermore, the image information richness parameter is calculated from information entropy. Higher information entropy indicates greater diversity of land cover types and richer details within the block, while lower entropy indicates less diverse information. Blocking according to this parameter ensures relatively uniform land cover characteristics within each block, facilitating targeted correction. The correction model is constructed based on the invariant probability obtained in step b3. Through the spectral correspondence of stable pixels, it calculates color correction parameters within each block, achieving initial color adjustment for each block and bringing the block colors of the multi-source remote sensing image to be processed closer to the surface reflectance product.

[0047] Step b5: Based on the distribution of each pixel position, the continuous correction parameter field of the multi-source remote sensing image to be processed is obtained by weighted bilinear interpolation, and color correction is completed pixel by pixel. The pixel position distribution is determined by the distance from each pixel to the center of each block.

[0048] Furthermore, the pixel location distribution is determined by the distance from the pixel to the center of its respective block: the closer the distance, the greater the influence of the correction parameters of that block on the pixel; the farther the distance, the higher the weight of the correction parameters of adjacent blocks. Through weighted bilinear interpolation, the discrete correction parameters of the blocks are converted into a continuous correction parameter field covering the entire multi-source remote sensing image to be processed, ensuring a smooth transition of correction parameters between adjacent blocks and avoiding color jumps. Finally, this parameter field is used to adjust the color of the multi-source remote sensing image to be processed pixel by pixel, making it consistent with the surface reflectance product globally, thus completely eliminating color differences between multi-source images.

[0049] In some alternative embodiments, Figure 3 This is a flowchart of a global and local color consistency processing method according to an embodiment of the present invention. The color consistency processing is a correction that combines global and local methods, including: Step c1: Use a global optimization strategy to uniformly calculate the color correction parameters of each multi-source remote sensing image to be processed.

[0050] Furthermore, the color characteristics of the overlapping areas of each pair of adjacent multi-source remote sensing images to be processed are first statistically analyzed. Assuming that the color differences between images follow a linear relationship, and with the goal of "consistent color tone in all overlapping areas of the adjusted images," the correction parameters of each multi-source remote sensing image to be processed, such as the luminance coefficient and contrast coefficient, are calculated globally using the least squares method. This avoids the cumulative errors that may occur from pairwise correction, ensuring that all multi-source remote sensing images to be processed have a unified color tone globally, and eliminating overall color differences caused by differences in sensor type and shooting time.

[0051] Step c2 involves using a local optimization strategy to eliminate residual color deviation in the overlapping areas of adjacent multi-source remote sensing images to be processed.

[0052] Furthermore, the local optimization strategy complements the global processing, eliminating subtle color differences that still exist in overlapping areas of adjacent multi-source remote sensing images. First, a minimum area covering all the images to be processed is determined. This area is then divided into multiple small blocks, and color information, such as average brightness and color distribution, is calculated for each block to construct a color distribution surface covering the entire area. For each pixel in the overlapping area, a reference color value is obtained from the color distribution surface using weighted bilinear interpolation based on its location. Then, a nonlinear adjustment is used to correct the pixel values ​​of the images to be processed towards the reference value. This approach focuses on local details, ensuring a natural transition between overlapping areas of adjacent multi-source remote sensing images and avoiding "striping" after stitching, thus further improving color consistency.

[0053] In some alternative embodiments, color consistency processing is a correction of the color reference library, including: The gray values ​​of the multi-source remote sensing image to be processed are mapped to the range of the gray mean and standard deviation of the target reference image through a linear transformation, as shown in the formula:

[0054] in, The grayscale values ​​of the multi-source remote sensing image to be processed. , These represent the mean and standard deviation of the grayscale values ​​of the multi-source remote sensing image to be processed. , These represent the mean and standard deviation of the grayscale value of the target reference image, respectively. The target reference image is extracted from a preset color reference library to obtain the processed grayscale value.

[0055] Furthermore, firstly, a target reference image is extracted from a pre-defined color reference library. This library stores standard color information for various scenarios. The target reference image, after screening, possesses balanced hues, reasonable grayscale distribution, and color characteristics suitable for power grid applications, serving as a benchmark for color correction of the multi-source remote sensing image to be processed. Next, grayscale statistical parameters of the multi-source remote sensing image to be processed and the target reference image are calculated. Specifically, the grayscale values ​​of all pixels in the multi-source remote sensing image to be processed are statistically analyzed to obtain its mean grayscale value and standard deviation. The former reflects the overall brightness of the image, while the latter reflects the dispersion of the image's grayscale values. Simultaneously, the mean grayscale value and standard deviation of the target reference image are obtained in the same manner. Finally, a linear transformation formula is used to map the grayscale values ​​of the multi-source remote sensing image to the range of the mean and standard deviation of the target reference image. The formula works as follows: First, it subtracts the mean gray value from the grayscale value of the multi-source remote sensing image to be processed, eliminating overall brightness deviation. Then, it scales the image using the ratio of the standard deviation between the target reference image and the multi-source remote sensing image to adjust the dispersion of the grayscale distribution. Finally, it adds the mean grayscale value of the target reference image, ensuring that the overall brightness of the processed image matches that of the reference image. Through this transformation, the color distribution of the multi-source remote sensing image to be processed is consistent with that of the target reference image, thus achieving color unification between different images.

[0056] Furthermore, Figure 4 This is a flowchart of color consistency processing based on a color reference library according to an embodiment of the present invention. The image to be processed queries the color reference library to extract reference color information. This library stores color parameters of standard or high-quality images as a benchmark for color alignment. Simultaneously, image segmentation and color information statistics are performed on the image to be processed. Segmentation divides the image into regions according to land cover type, and the mean and variance of color in each region are statistically analyzed to understand its own color distribution. Then, combining the reference color information and the color statistical results of the image to be processed, a color weight map is established and color consistency processing is performed. The color weight map assigns and adjusts weights to different regions to make the processing more accurate. Color consistency processing uses algorithms to align the colors of the image to be processed with the reference colors, eliminating color differences caused by sensors and shooting conditions in multi-source images. Finally, a color-unified result image is output, laying a solid foundation for subsequent multi-source remote sensing image stitching. This ensures that the stitched image is not only geometrically seamless but also naturally connected in color, meeting the requirements of image visual and data consistency in scenarios such as power grid monitoring. In conjunction with processes such as radiometric correction, it comprehensively improves image quality and ensures stitching effect.

[0057] Step S102: For the preprocessed multi-source remote sensing image set, determine the intersecting n-sided polygons between adjacent preprocessed multi-source remote sensing images to obtain several intersecting n-sided polygons, where n is a positive integer greater than or equal to 3.

[0058] Specifically, step S102 includes: Step d1: Parse the vertex coordinates of the effective range from the metadata of the current preprocessed multi-source remote sensing image to form a polygon describing the effective range, and obtain the boundary coordinates of the effective range of the multi-source remote sensing image to be processed.

[0059] Furthermore, metadata is supplementary data that stores geographic information about the imagery, including information such as the latitude and longitude or planar coordinates of the image coverage. The vertex coordinates of the effective extent are parsed from the metadata, and these vertices are connected sequentially to form a closed polygon. This polygon represents the geometric boundary describing the effective extent of the multi-source remote sensing image to be processed. Through this step, the effective extent of the image is transformed into a computable geometric shape, obtaining specific boundary coordinate values, providing numerical basis for subsequent spatial relationship analysis.

[0060] Step d2: Based on the effective range boundary coordinates, use a spatial indexing algorithm to retrieve adjacent preprocessed multi-source remote sensing images that spatially overlap with the effective range of the current preprocessed multi-source remote sensing image.

[0061] Specifically, the spatial indexing algorithm is a highly efficient spatial data retrieval technique. It quickly locates other images that overlap with the effective range of the current image by indexing the boundary coordinates of the effective range of all preprocessed multi-source remote sensing images. The algorithm compares the spatial positional relationships between the polygons of the current image and those of other images, filtering out images with intersecting or contained boundaries, and marking them as adjacent preprocessed multi-source remote sensing images.

[0062] Step d3: Use polygon Boolean intersection operation to calculate the effective range polygons of the current preprocessed multi-source remote sensing image and the adjacent preprocessed multi-source remote sensing images, and solve to obtain a closed intersecting n-sided polygon.

[0063] Furthermore, the polygon Boolean intersection operation is used to calculate the overlapping area of ​​two polygons. Through geometric algorithms, the effective range polygons of the current preprocessed multi-source remote sensing image and the adjacent preprocessed multi-source remote sensing image are calculated to obtain the area shared by the two. This area is a closed simple polygon with non-intersecting edges, i.e., an intersecting n-sided polygon, where n is a positive integer greater than or equal to 3. This accurately extracts the overlapping parts of the adjacent images and clarifies the stitching transition area that needs to be processed.

[0064] Step d4: Traverse each preprocessed multi-source remote sensing image in the preprocessed multi-source remote sensing image set, and remove intersecting n-gons with an area smaller than a preset threshold to obtain several intersecting n-gons.

[0065] Furthermore, the preset threshold is set according to the actual splicing accuracy requirements. When the splicing accuracy requirements are: When the geometric error in the overlapping area of ​​the image after stitching does not exceed 5 pixels and the color transition is natural without obvious discontinuities, the preset threshold is set to "the area of ​​the intersecting n-sided polygon corresponds to an actual ground area of ​​not less than 1000 square meters". Assuming the image resolution is 1 meter, the pixel area of ​​the intersecting n-sided polygon is not less than 1000 pixels.

[0066] If the area of ​​the intersecting n-sided polygon is too small, the central axis and bisectors generated subsequently will lack sufficient adjustment space due to their narrow range, which may easily lead to geometric misalignment or harsh color transitions during splicing, failing to meet the accuracy requirements. However, when the area of ​​the intersecting n-sided polygon is not less than the corresponding threshold, it can provide sufficient area for the generation of the central axis and bisectors, thereby ensuring splicing accuracy.

[0067] Traverse all preprocessed multi-source remote sensing images, calculate the area of ​​the intersecting n-sided polygons of each image with its neighboring images, and consider the intersecting n-sided polygons with an area less than a preset threshold as invalid overlaps, such as slight edge overlaps, which have no practical splicing significance and are then discarded.

[0068] The retained intersecting n-gons are all valid overlapping areas with the required area, which can be used to generate the central axis and bisectors in the subsequent process, ensuring that the splicing process focuses on the overlapping parts with practical significance and improving the overall processing efficiency.

[0069] Step S103: Determine the central axis of each of the intersecting n-gons, and generate the bisector between adjacent preprocessed multi-source remote sensing images based on the central axis.

[0070] Specifically, step S103 includes: Step e1: Identify the intersection points of the effective range boundaries of adjacent preprocessed multi-source remote sensing images on the intersecting n-sided polygon, and select the two intersection points that are farthest apart as the start point and the end point, respectively.

[0071] Furthermore, the intersecting n-gon is a closed region formed by the overlap of the effective ranges of adjacent preprocessed multi-source remote sensing images, and its boundary is jointly formed by the effective range boundaries of the two images. Geometric calculations are used to find all intersection points of the effective range boundaries of the two images on the intersecting n-gon; these intersection points are the vertices of the overlapping region. The two points furthest apart from these intersection points are selected as the start and end points for subsequent central axis calculations, ensuring that the central axis runs through the entire intersecting n-gon.

[0072] Step e2: Arrange all the vertices of the intersecting n-sided polygons in order to form a vertex sequence, and extract the line segment between the starting point and the ending point at the corresponding positions in the vertex sequence as the central axis between adjacent images.

[0073] Furthermore, the vertices of the intersecting n-gons are arranged sequentially in a clockwise or counterclockwise direction, forming a continuous sequence of inflection points, each representing a change in boundary direction. The positions corresponding to the start and end points are found from the inflection point sequence, and a line segment is extracted between these two positions. This line segment is approximately equidistant from the two boundaries of the intersecting n-gons, and thus forms the central axis of the overlapping region of adjacent preprocessed multi-source remote sensing images.

[0074] Step e3: Based on the image segmentation algorithm, identify the artificial building regions within the intersecting n-sided polygons and generate a binary mask.

[0075] Furthermore, the image segmentation algorithm distinguishes areas with artificial buildings from other areas by analyzing the features of ground features within intersecting n-gons. The generated binary mask is a black and white image in which areas with artificial buildings are marked with a specific value of 1, and other areas are marked with 0, thus clearly defining the areas that need to be avoided.

[0076] Step e4: If the central axis passes through a building area, shift the central axis away from the building by a preset distance.

[0077] Furthermore, by comparing the central axis with the binary mask, it is determined whether the central axis passes through areas marked as artificial buildings. If there is a crossing, the central axis is shifted away from the building by a preset distance to ensure that the central axis avoids artificial buildings and prevents the subsequently generated bisector from passing through extra ground features, which could cause misalignment in the stitching. The preset distance is set according to the size of the building and the stitching accuracy requirements.

[0078] The preset distance is set using a computer-executable method of "image resolution quantization + dynamic calculation of building pixel scale". It reads the metadata of the preprocessed multi-source remote sensing images, extracts the image spatial resolution parameters and uses them as the basis for calculation. The corresponding baseline offset pixel count is initialized according to the resolution to ensure consistent baseline safety distances for images of different resolutions. A linear relationship based on an empirical formula is used to correlate the resolution with the initial baseline offset pixel count. The specific formula is as follows:

[0079] Where N is the initial reference offset pixel count, R is the spatial resolution of the preprocessed multi-source remote sensing image, and k and b are empirical coefficients.

[0080] Through extensive experimental verification using multi-resolution images, it was determined that when pursuing a safe distance between the central axis and the edge of the building after stitching, it is necessary to effectively avoid buildings without excessively wasting overlapping areas. .

[0081] For example, when the image resolution Initial reference offset pixels per meter ; When image resolution Initial reference offset pixels per meter At this point, the number of pixels can be rounded (e.g., rounded to 3) to suit the integer characteristics of pixels.

[0082] This linear relationship ensures that the initial reference offset pixel count for images at different resolutions is proportional to the resolution, guaranteeing that the actual safe distance (offset pixel count × resolution) between the central axis and the building edge remains relatively consistent across different resolutions. This provides a reasonable reference starting point for subsequent dynamic adjustments based on the building's pixel scale. Next, pixel statistics are performed on the artificial building areas in the binary mask. The total number of pixels occupied by a single building in the mask is calculated using a connected component analysis algorithm. Based on the range of the total number of pixels, the number of reference offset pixels is adjusted accordingly. When the total number of pixels is large, the number of reference offset pixels is increased; when the total number of pixels is small, the number of reference offset pixels is decreased; when the total number of pixels is in the middle range, the number of reference offset pixels remains unchanged. Finally, the final determined number of offset pixels is converted into the offset distance in image coordinates. Based on the geometric direction of the central axis, the offset direction away from the building is determined by vector calculation, thus completing the automatic setting of the preset distance.

[0083] Step e5 involves curve fitting and path simplification of the offset central axis to generate the bisector between adjacent preprocessed multi-source remote sensing images.

[0084] Furthermore, curve fitting uses mathematical methods to smooth the offset central axis, eliminating sharp angles and abrupt changes in the broken line segments, making the path more natural. Path simplification removes redundant vertices on the central axis, retains key turning points, reduces data volume, and ensures path continuity. The smoothed curve obtained after processing serves as the bisector between adjacent preprocessed multi-source remote sensing images, acting as the boundary for subsequent cropping and stitching to ensure a natural transition in overlapping areas.

[0085] Step S104: Based on the bisecting line, crop each of the preprocessed multi-source remote sensing images to obtain the target stitching line.

[0086] Specifically, step S104 includes: Step f1: Use the effective range of the current preprocessed multi-source remote sensing image as the initial cropping region.

[0087] Furthermore, the effective extent of the preprocessed multi-source remote sensing image is the actual coverage area determined after geometric correction, containing valid ground feature information and geographic coordinates within its boundaries. Using the effective extent as the initial cropping area means that the cropping operation begins with the complete effective extent of the image, and will be gradually adjusted according to the bisector, providing a foundation for accurately delineating the image's specific regions.

[0088] Step f2: Integrate the bisecting lines between the current preprocessed multi-source remote sensing image and the adjacent preprocessed multi-source remote sensing images to construct a bisecting line network.

[0089] Furthermore, the bisector is the central dividing line within the overlapping region of adjacent preprocessed multi-source remote sensing images, generated by optimizing the central axis of intersecting n-gons. Integrating the bisectors between the current image and all adjacent images forms an interconnected bisector network, providing a unified segmentation basis for subsequent batch cropping and ensuring that the cropping boundaries of adjacent images correspond to each other.

[0090] Step f3: Use the bisecting line network to crop the preprocessed multi-source remote sensing image sequentially, preserving the centroid region.

[0091] Furthermore, the centroid region is the core area near the geometric center of the preprocessed multi-source remote sensing image, and the land cover information within this region best represents the original features of the image. During cropping, the initial cropping area is segmented sequentially according to each bisector of the bisecting line network. Each cropping retains the portion close to the current image centroid and removes overlapping areas far from the centroid, making the remaining area more closely match the core coverage of the image.

[0092] Step f4: Traverse each preprocessed multi-source remote sensing image in the preprocessed multi-source remote sensing image set to generate several Voronoi polygons.

[0093] Furthermore, steps f1 to f3 are repeated for each preprocessed multi-source remote sensing image in the set. First, the effective range of each image is used as the initial region, then it is cropped using a network of bisecting lines with adjacent images, preserving the centroid region. Finally, the closed region formed after cropping each image is a Voronoi polygon, characterized by the fact that the distance from any point within the polygon to the centroid of the corresponding image is less than the distance to the centroid of any other image, and all polygons collectively cover the entire processing area.

[0094] Step f5: Check the several Voronoi polygons to determine if there are any conflicts between them. If there are conflicts between the several Voronoi polygons, adjust the bisecting network and re-trim it until there are no conflicts between the several Voronoi polygons.

[0095] Furthermore, a conflict refers to an overlap or gap between polygons. By traversing all Voronoi polygons using a spatial geometry inspection tool, if a conflict is found, the corresponding bisector positions need to be adjusted, the bisector network reconstructed, and a clipping process executed until all polygons seamlessly connect, do not overlap, and completely cover the entire area.

[0096] Step f6 defines the common boundary of several Voronoi polygons as the target splicing line.

[0097] Furthermore, the intersection of the boundaries of any two Voronoi polygons is calculated one by one. If the intersection exists and the area is greater than zero, it is determined to be an overlapping conflict. At the same time, the minimum bounding rectangle of the entire region to be stitched is traversed to check if there is a region not covered by any Voronoi polygon. If so, it is determined to be a gap conflict. The adjustment adopts a fully automatic algorithm, with the following rules: To address overlapping conflicts, the computer calculates the geometric center of the overlapping region and moves the bisector of the overlapping region to the side of the polygon that is farther from the center. The magnitude of the movement is dynamically determined based on the scale of the overlapping region. To address gap conflicts, the computer identifies the Voronoi polygons surrounding the gap and fine-tunes the corresponding bisectors towards the gap, extending the boundaries of adjacent polygons to cover the gap.

[0098] After each adjustment, the computer reconstructs the bisecting network and performs clipping, generating new Voronoi polygons. The collision detection process is repeated until all polygon boundaries are non-intersecting and completely cover the entire area.

[0099] When all Voronoi polygons are conflict-free, the shared boundaries between adjacent polygons are called common boundaries. These boundaries are formed by line segments in the bisecting line network and are formally defined as target stitching lines. As the final dividing line for image stitching, the target stitching line ensures that each preprocessed multi-source remote sensing image participates in stitching only within its own Voronoi polygons, avoiding duplication or omission of overlapping areas and providing precise boundaries for seamless stitching.

[0100] Step S105: Based on the target stitching line, each of the preprocessed multi-source remote sensing images is cropped, and the cropped images are stitched together to obtain a seamless target stitched image.

[0101] Specifically, the target stitching lines are determined by the common boundaries of the Voronoi polygons, with each stitching line corresponding to the segmentation boundary of adjacent preprocessed multi-source remote sensing images. During cropping, following the path of the stitching lines, the portion of each preprocessed multi-source remote sensing image within its respective Voronoi polygon is retained, while areas outside the polygons are removed. This ensures that each image contributes only its optimal region to the stitching, avoiding redundancy or conflicts in overlapping parts. The stitching process aligns the cropped data according to spatial location based on the geographic coordinates of the images, allowing adjacent images to connect naturally at the stitching lines. Since color consistency processing was completed during preprocessing, and the stitching lines avoid areas prone to misalignment such as man-made structures, the stitched images are seamlessly aligned geometrically and exhibit natural and unified color transitions, ultimately forming a seamless target stitched image covering the entire area. This image has no obvious stitching marks, clearly presents ground feature details, and meets the needs of large-scale, high-precision imagery for scenarios such as UHV transmission line monitoring and geological disaster early warning in power grid applications.

[0102] In some optional embodiments, the target seamless mosaic image is processed into sectional images to generate sectional images that conform to a preset coordinate system and file organization standard. The preset coordinate system includes the CGCS2000 national geodetic coordinate system and the WGS84 coordinate system. The target seamless mosaic image supports latitude and longitude projection and Gauss-Krüger projection.

[0103] Specifically, when performing tiled processing on seamlessly stitched images of a target, it is first necessary to clarify the specific requirements for the preset coordinate system and projection method. The preset coordinate systems include the CGCS2000 national geodetic coordinate system and the WGS84 coordinate system. The CGCS2000 coordinate system is suitable for domestic power grid planning, line monitoring and other businesses; the WGS84 coordinate system is an internationally used geocentric coordinate system, which is convenient for image applications in cross-regional or international cooperation scenarios.

[0104] The target seamless image stitching supports latitude and longitude projection, which is a projection method that directly represents the image position with longitude and latitude. It is suitable for macro-level browsing of large areas and cross-regional stitching. Gauss-Kruger projection, on the other hand, converts spherical coordinates into planar coordinates through zone projection, which can effectively reduce the deformation in mid-latitude regions and is suitable for high-precision application scenarios such as power grid line design and engineering surveying.

[0105] The specific process of segmentation processing is as follows: Based on a pre-defined coordinate system and file organization standards, the large-scale seamless imagery of the target area is divided into fixed map sheet sizes. During segmentation, it is essential to ensure that the boundaries of each image sheet are aligned with the grid lines of the coordinate system, and that map sheet numbering, file naming, and other aspects follow a unified file organization standard. For images supporting multiple projections, latitude and longitude projection or Gauss-Krüger projection can be selected for segmentation according to business needs. The resulting segmented images satisfy both the requirements of large-scale imagery for macro-level power grid monitoring and the high-precision planar coordinate requirements of refined line design, ultimately forming standardized segmented imagery products that facilitate storage, management, and business applications.

[0106] The graph theory-based seamless image stitching method provided in this embodiment first acquires geometrically corrected orthophotos as a preprocessed set of multi-source remote sensing images to ensure that the images have accurate geographic coordinate references and eliminate the geometric offset of the original images. At the same time, the images are down-digitized and color consistent with each other. This not only solves the problem that high-digit images are difficult to display on PCs or web platforms, but also eliminates color deviations caused by different satellite sensors and different shooting conditions. It effectively avoids stitching errors caused by inconsistent data formats or color differences, while meeting the special requirements of image geometric accuracy and color consistency in power grid applications.

[0107] Secondly, by determining the intersecting n-gons between adjacent preprocessed multi-source remote sensing images, the boundaries of overlapping areas between images are accurately identified, clarifying the overlapping parts that need to be focused on during the stitching process. This provides accurate geometric range and boundary conditions for subsequent central axis calculation and bisector generation, effectively eliminating interference from non-overlapping areas, reducing invalid calculations, and improving processing efficiency. At the same time, it ensures that the subsequently generated stitching lines are completely located within the overlapping area, avoiding stitching misalignment or gaps caused by blurred boundaries. It is particularly suitable for processing complex overlapping relationships of large-scale multi-source images, providing accurate spatial positioning basis for achieving seamless global stitching.

[0108] Subsequently, by determining the central axis of each intersecting n-gon, and generating bisectors between adjacent images based on the central axis, the bisectors are ensured to always be distributed along the central path of the overlapping area by utilizing the property that the distance from the polygon's central axis to the boundary is equal. This mathematically guarantees the rationality of the stitching line. At the same time, combined with the image's ground feature characteristics, the central axis automatically avoids key ground features such as high-rise buildings and towers during the generation process, avoiding the stitching misalignment caused by traditional stitching lines passing through redundant ground features. The generated bisectors have both continuity and smoothness, providing a stable reference boundary for subsequent cropping and stitching, significantly improving the geometric accuracy and visual effect of the stitching result, and meeting the requirements for image detail integrity in power grid inspection.

[0109] Then, by cropping each image to be processed based on the bisecting lines between adjacent images, a target mosaicking line composed of Voronoi polygon boundaries is generated. This achieves accurate division of the mosaicking area for each preprocessed multi-source remote sensing image, ensuring that all Voronoi polygons do not overlap and completely cover the entire area. This avoids the reuse or omission of overlapping areas, allowing each preprocessed multi-source remote sensing image to contribute only its best area for mosaicking. At the same time, the target mosaicking line inherits the ground feature avoidance characteristics of the bisecting lines, further ensuring the rationality of the mosaicking boundary. This provides a clear and executable basis for subsequent mosaicking operations, effectively improving the automation level and processing efficiency of large-scale image mosaicking, and adapting to the large-scale production needs of "weekly updates of meter-level images and monthly updates of sub-meter-level images" in power grid operations.

[0110] Finally, the preprocessed images were cropped based on the target stitching line, and then stitched together. The geometrically corrected and color-unified image fragments were integrated into a complete seamless target image according to geographic coordinates, fully leveraging the advantages of the high-precision data source and optimized stitching line in the previous steps. The stitched result not only achieved seamless alignment in geometric position but also maintained natural consistency in color transition, meeting the needs of large-scale, high-precision imagery for scenarios such as UHV dense channel inspection, flood control and typhoon prevention, and geological disaster early warning in power grid applications. At the same time, the seamless imagery can be further processed into sectional images according to power grid standards, providing unified and reliable spatial data support for subsequent line planning, hazard identification, emergency decision-making, and other operations, promoting the deep application of satellite remote sensing data in the digital transformation of the power grid.

[0111] By implementing this invention, the problem of low data reuse rate and difficulty in cross-scenario application caused by the lack of a standardized processing framework for the entire process of remote sensing image stitching technology for power grid applications is solved.

[0112] In Embodiment 2, this invention provides a graph-based seamless remote sensing image stitching system, which implements Embodiment 1 and the preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the systems described in the following embodiments are preferably implemented in software, hardware implementations, or combinations of software and hardware, are also possible and contemplated.

[0113] This embodiment provides a graph theory-based seamless image stitching system for remote sensing, such as... Figure 5 As shown, it includes: The module comprises a preprocessing module 501, a determining module 502, a generating module 503, a trimming module 504, and a splicing module 505, wherein: The preprocessing module 501 is used to acquire a set of multi-source remote sensing images to be processed, and to preprocess each multi-source remote sensing image in the set of multi-source remote sensing images to be processed to obtain a preprocessed set of multi-source remote sensing images. The preprocessing includes downsampling and color consistency processing. The multi-source remote sensing images to be processed are orthophotos after geometric correction. The determining module 502 is used to determine the intersecting n-sided polygons between adjacent preprocessed multi-source remote sensing images for the preprocessed multi-source remote sensing image set, and obtain several intersecting n-sided polygons, where n is a positive integer greater than or equal to 3; The generation module 503 is used to determine the central axis of each of the intersecting n-gons and generate the bisector between adjacent preprocessed multi-source remote sensing images based on the central axis. The cropping module 504 is used to crop each of the preprocessed multi-source remote sensing images based on the bisector to obtain the target stitching line; The stitching module 505 is used to crop each of the preprocessed multi-source remote sensing images based on the target stitching line, and to stitch the cropped images together to obtain a target seamless stitched image.

[0114] The technical effects of this invention include: 1. The graph theory-based seamless remote sensing image stitching system provided in this embodiment acquires geometrically corrected orthophotos as a set of multi-source remote sensing images to be processed, ensuring that the images have accurate geographic coordinate references, eliminating geometric offsets of the original images, and performing down-biting and color consistency processing on the images. This solves the problem that high-bit images are difficult to display on PCs or web platforms, and eliminates color deviations caused by different satellite sensors and shooting conditions. It effectively avoids stitching errors caused by inconsistent data formats or color differences, while meeting the special requirements for image geometric accuracy and color consistency in power grid applications.

[0115] 2. The graph theory-based seamless remote sensing image stitching system provided in this embodiment accurately identifies the boundaries of overlapping areas between adjacent multi-source remote sensing images by determining the intersecting n-gons between them. This clarifies the overlapping parts that need to be focused on during the stitching process, providing accurate geometric range and boundary conditions for subsequent central axis calculation and bisector generation. It can effectively eliminate interference from non-overlapping areas, reduce invalid calculations, and improve processing efficiency. At the same time, it ensures that the subsequently generated stitching lines are completely located within the overlapping area, avoiding stitching misalignment or gaps caused by blurred boundaries. It is especially suitable for processing complex overlapping relationships of large-scale multi-source images, providing accurate spatial positioning basis for achieving global seamless stitching.

[0116] 3. The graph theory-based seamless remote sensing image stitching system provided in this embodiment determines the central axis of each intersecting n-gon and generates bisectors between adjacent images based on the central axis. Utilizing the property that the distance from the polygon's central axis to the boundary is equal, it ensures that the bisectors are always distributed along the central path of the overlapping area, thus guaranteeing the rationality of the stitching lines from a mathematical perspective. At the same time, combined with the image's ground feature characteristics, the central axis automatically avoids key ground features such as high-rise buildings and towers during the generation process, avoiding stitching misalignment caused by traditional stitching lines passing through redundant ground features. The generated bisectors have both continuity and smoothness, providing a stable reference boundary for subsequent cropping and stitching, significantly improving the geometric accuracy and visual effect of the stitching results, and meeting the requirements for image detail integrity in power grid inspection.

[0117] 4. The graph theory-based seamless remote sensing image stitching system provided in this embodiment clips each image to be processed based on the bisecting lines between adjacent images, generating a target stitching line composed of Voronoi polygon boundaries. This achieves accurate division of the stitching area for each multi-source remote sensing image to be processed, ensuring that all Voronoi polygons do not overlap and completely cover the entire area, avoiding the reuse or omission of overlapping areas, and allowing each multi-source remote sensing image to contribute only its best area for stitching. At the same time, the target stitching line inherits the ground feature avoidance characteristics of the bisecting lines, further ensuring the rationality of the stitching boundary, providing a clear and executable basis for subsequent stitching operations, effectively improving the automation level and processing efficiency of large-scale image stitching, and adapting to the large-scale production needs of "weekly updates of meter-level images and monthly updates of sub-meter-level images" in power grid operations.

[0118] 5. The graph theory-based seamless remote sensing image stitching system provided in this embodiment crops the preprocessed image based on the target stitching line and then stitches the cropped image together. This integrates geometrically corrected and color-unified image segments into a complete seamless image based on geographic coordinates, fully leveraging the advantages of the high-precision data source and optimized stitching line in the previous steps. The stitching result not only achieves seamless alignment in geometric position but also maintains natural consistency in color transition, meeting the needs of large-scale, high-precision imagery for scenarios such as UHV dense channel inspections, flood and typhoon prevention, and geological disaster early warning in power grid applications. Furthermore, the seamless image can be further processed into sectional images according to power grid standards, providing unified and reliable spatial data support for subsequent line planning, hazard identification, and emergency decision-making, thus promoting the deep application of satellite remote sensing data in the digital transformation of the power grid.

[0119] Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0120] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A graph-theory-based method for seamless stitching of remote sensing images, characterized in that, The method includes: A set of multi-source remote sensing images to be processed is obtained, and each multi-source remote sensing image in the set is preprocessed to obtain a set of preprocessed multi-source remote sensing images. The multi-source remote sensing images to be processed are orthophotos after geometric correction. The preprocessing includes performing down-biting processing first, and then performing color consistency processing. For the preprocessed multi-source remote sensing image set, the intersecting n-sided polygons between adjacent preprocessed multi-source remote sensing images are determined, resulting in several intersecting n-sided polygons, where n is a positive integer greater than or equal to 3. Determine the central axis of each of the intersecting n-gons, and generate the bisectors between adjacent preprocessed multi-source remote sensing images based on the central axes, including: Identify the intersection points of the effective range boundaries of adjacent preprocessed multi-source remote sensing images on intersecting n-sided polygons, and select the two farthest intersection points as the start and end points respectively; Arrange the vertices of the intersecting n-sided polygons in order to form a sequence of vertices, and take the broken line segment between the starting point and the ending point as the central axis; Based on image segmentation, identify artificial building regions within intersecting n-sided polygons and generate corresponding binary masks; When the central axis passes through a building area, the central axis will be offset away from the building by a preset distance; Curve fitting and path simplification are performed on the offset central axis to generate the bisector; Based on the bisecting line, each of the preprocessed multi-source remote sensing images is cropped to obtain the target stitching line; Based on the target stitching line, each of the preprocessed multi-source remote sensing images is cropped, and the cropped images are stitched together to obtain a seamless target stitched image.

2. The graph-based seamless image stitching method according to claim 1, characterized in that: The downsizing process includes: The multi-source remote sensing images to be processed are stretched using nonlinear Gamma stretching. Based on the normalized vegetation index and the normalized water index, the vegetation area and water area in each multi-source remote sensing image to be processed are enhanced respectively. The stretched and enhanced multi-source remote sensing images to be processed are converted into 8-bit multi-source remote sensing images to be processed.

3. The graph-based seamless image stitching method according to claim 1, characterized in that: The color consistency processing is a hybrid radiometric correction, including: Candidate reference images are determined based on the geographic extent, acquisition time, and spatial resolution of the multi-source remote sensing images to be processed; Radiometric calibration and atmospheric correction are performed on the candidate reference images to obtain the surface reflectance products of the candidate reference images; The multi-source remote sensing image to be processed is resampled until the resolution of the multi-source remote sensing image to be processed and the surface reflectance product are consistent. The invariance probability of corresponding pixels between the multi-source remote sensing image to be processed and the surface reflectance product is calculated based on the IR-MAD algorithm. Based on the rich parameters of image information, the multi-source remote sensing image to be processed is divided into blocks, and each block is corrected based on the correction model constructed based on invariant probability. Based on the distance distribution from each pixel to its block center, a continuous correction parameter field is generated by weighted bilinear interpolation, and pixel-by-pixel color correction is performed on the multi-source remote sensing image to be processed according to the correction parameter field.

4. The graph-theory-based seamless image stitching method according to claim 1, characterized in that: The color consistency processing is a correction that combines global and local methods, including: A global optimization strategy is used to uniformly calculate the color correction parameters of each multi-source remote sensing image to be processed; A local optimization strategy is used to eliminate residual color deviation in the overlapping areas of adjacent multi-source remote sensing images to be processed.

5. The graph-theory-based seamless image stitching method according to claim 1, characterized in that: The color consistency processing is a correction based on a color reference library, including: The grayscale values ​​of the multi-source remote sensing images to be processed are mapped to the grayscale mean and standard deviation range of the target reference image through linear transformation. The target reference image is extracted from a preset color reference library.

6. The graph-theory-based seamless image stitching method according to claim 1, characterized in that: Determining the intersecting n-gons between adjacent preprocessed multi-source remote sensing images includes: Geometric correction is performed on the preprocessed multi-source remote sensing images to determine their effective range; The vertex coordinates of the effective range are parsed from the metadata of the preprocessed multi-source remote sensing image to form a polygon describing the effective range, and the boundary coordinates of the effective range are obtained. Based on the effective range boundary coordinates, a spatial indexing algorithm is used to retrieve adjacent preprocessed multi-source remote sensing images whose effective ranges overlap with those of the preprocessed multi-source remote sensing images. Polygon Boolean intersection operation is used to calculate the effective range polygons of the preprocessed multi-source remote sensing image and adjacent preprocessed multi-source remote sensing images, and the closed intersecting n-sided polygon is obtained by solving. Traverse each preprocessed multi-source remote sensing image in the preprocessed multi-source remote sensing image set, and remove intersecting n-gons with an area smaller than a preset threshold to obtain several intersecting n-gons.

7. The graph-theory-based seamless image stitching method according to claim 1, characterized in that: The preset distance setting includes: Based on the spatial resolution of the preprocessed multi-source remote sensing images, a baseline offset is determined. The reference offset is dynamically adjusted based on the pixel size of the artificial building area in the binary mask. The larger the pixel size of the building area, the greater the adjustment range of the reference offset. The adjusted offset is converted into an actual distance value, and the offset direction is determined by combining it with the geometric orientation of the central axis to complete the setting of the preset distance.

8. The graph-theory-based seamless image stitching method according to claim 1, characterized in that: The target splicing lines include: Use the effective range of the current preprocessed multi-source remote sensing image as the initial cropping region; Integrate the bisecting lines between the current preprocessed multi-source remote sensing images and adjacent preprocessed multi-source remote sensing images to construct a bisecting line network; The preprocessed multi-source remote sensing images are cropped sequentially using a bisecting line network, preserving the centroid region; Traverse each preprocessed multi-source remote sensing image in the preprocessed multi-source remote sensing image set to generate several Voronoi polygons; Check several Voronoi polygons to determine if there are any conflicts between them. If there are conflicts, adjust the bisecting network and re-trim until there are no conflicts between the several Voronoi polygons. Define the common boundary of several Voronoi polygons as the target splicing line.

9. A graph-based seamless remote sensing image stitching system, used to perform the graph-based seamless remote sensing image stitching method according to any one of claims 1-8, characterized in that, include: The module includes a preprocessing module, a determining module, a generating module, a cropping module, and a stitching module, among which: The preprocessing module is used to acquire a set of multi-source remote sensing images to be processed, and to preprocess each multi-source remote sensing image in the set of multi-source remote sensing images to be processed to obtain a preprocessed set of multi-source remote sensing images. The preprocessing includes downsampling and color consistency processing. The multi-source remote sensing images to be processed are orthophotos after geometric correction. The determining module is used to determine the intersecting n-sided polygons between adjacent preprocessed multi-source remote sensing images for the preprocessed multi-source remote sensing image set, and obtain several intersecting n-sided polygons, where n is a positive integer greater than or equal to 3; The generation module is used to determine the central axis of each of the intersecting n-gons, and generate the bisector between adjacent preprocessed multi-source remote sensing images based on the central axis. The cropping module is used to crop each of the preprocessed multi-source remote sensing images based on the bisector to obtain the target stitching line; The stitching module is used to crop each of the preprocessed multi-source remote sensing images based on the target stitching line, and to stitch the cropped images together to obtain a target seamless stitched image.

Images