Remote sensing redundant data identification method based on multi-view geometry theory and storage medium

Abstract

The application provides a remote sensing redundant data recognition method based on multi-view geometry theory and a storage medium, relates to the field of image processing, and solves the technical problems that the prior art is insufficient in perceiving the geometric nature of data redundancy, cannot effectively quantify and evaluate the perspective redundancy, and the existing redundant detection process is disconnected from the model training process. The method comprises the following steps: obtaining an original aerial image sequence and preprocessing the original aerial image sequence to obtain a preprocessed image sequence; processing the preprocessed image sequence based on a motion recovery structure technology to obtain camera poses and three-dimensional point clouds of the images and quantify the geometric redundancy between the images; extracting visual features of the images, combining the geometric redundancy, and generating a comprehensive redundancy score; identifying a redundant image cluster in the image sequence according to the comprehensive redundancy score, and retaining an optimal image from each cluster according to a preset rule to obtain an optimized image set.

Description

Technical Field

[0001] This invention belongs to the field of image processing, specifically a method for identifying redundant remote sensing data based on multi-view geometry theory. Background Technology

[0002] With the rapid development of low-altitude remote sensing technology, especially the widespread application of consumer and professional drone platforms, aerial photography for geospatial data acquisition has become a standard technique in fields such as digital twins, smart cities, 3D modeling, and precision agriculture. To ensure the accuracy and completeness of subsequent 3D reconstruction, low-altitude aerial photography operations typically employ highly overlapping flight path designs. This directly leads to significant data redundancy when the acquired image sequences are used directly for deep learning model training. This redundancy manifests primarily as repeated observations of the same ground features from similar or adjacent perspectives, exhibiting high repetition at the pixel-level features, visual content, and even 3D geometric levels. During model training, this redundancy results in wasted computational resources, low training efficiency, and potential overfitting, severely limiting the processing efficiency and application value of low-altitude remote sensing data.

[0003] Currently, the technical solutions for handling redundancy in low-altitude remote sensing data generally suffer from the following three core bottlenecks: First, traditional methods mainly rely on the similarity comparison of two-dimensional image features, such as features extracted based on color histograms and machine vision algorithms, to determine redundancy. This method cannot penetrate apparent variations such as illumination changes, scale differences, and viewpoint rotation, and is unable to address the essential redundancy of the same ground feature in three-dimensional space.

[0004] Secondly, in low-altitude photogrammetry, the value of images lies not only in their two-dimensional content but also in the stereoscopic information provided by their unique perspective. Current methods cannot quantitatively analyze camera pose, making it difficult to distinguish between complementary images essential for 3D reconstruction and redundant images present in deep learning models. This deficiency often leads to the accidental deletion of crucial image data during deep learning model training, while retaining a large number of redundant images with overlapping perspectives and minimal contribution.

[0005] Third, existing redundancy detection processes are inefficient and disconnected from the main task of model training. Many redundancy detection schemes are treated as independent pre-modules, and their judgment criteria are inconsistent with those used in subsequent model training. This can lead to uneven data distribution in the cleaned dataset during later training stages, causing the model to favor the majority class and resulting in extremely poor performance in recognizing the minority class. Summary of the Invention

[0006] This application provides a remote sensing redundancy data identification method based on multi-view geometry theory, which solves the problems of insufficient perception of the geometric nature of data redundancy in existing technologies, inability to effectively quantify and evaluate view redundancy, and the technical problem of the disconnect between existing redundancy detection process and model training process.

[0007] To achieve the above objectives, this application adopts the following technical solution: Firstly, a method for identifying redundant remote sensing data based on multi-view geometry theory is provided, including: The original aerial image sequence is acquired and preprocessed to obtain the preprocessed image sequence. The preprocessed image sequence is processed using motion recovery structure technology to obtain the camera pose and 3D point cloud of the image, and the geometric redundancy between the images is quantified. Visual features of the image are extracted and combined with the geometric redundancy to generate a comprehensive redundancy score. Based on the redundancy comprehensive score, redundant image clusters in the image sequence are identified, and the best image is retained from each cluster according to preset rules to obtain an optimized image set.

[0008] Based on the above technical solutions, the remote sensing redundancy data identification method based on multi-view geometry theory provided in this application starts from the essence of three-dimensional geometry and uses motion reconstruction structure technology to obtain the accurate camera pose and three-dimensional point cloud of the image. This overcomes the misjudgment problem caused by changes in illumination, scale, and viewpoint in traditional two-dimensional feature methods, significantly improving the accuracy of redundancy identification. Then, by quantifying geometric redundancy and combining it with visual features to generate a comprehensive score, the method achieves multi-dimensional redundancy assessment. In addition, the redundancy cluster identification and optimal image preservation mechanism based on the comprehensive score not only automates the data cleaning process and greatly reduces manual intervention and computational resource consumption, but also ensures that the optimized image set retains key information while improving the efficiency and quality of subsequent three-dimensional reconstruction or model training.

[0009] Furthermore, the preprocessing steps for the original aerial image sequence include: Establish the correspondence between images and POS data from the positioning and attitude determination system; Automatic exposure compensation or histogram equalization is performed on the original aerial image to obtain the first pre-processed image after radiometric adjustment. Using POS data, the initial extrinsic parameters of all the first preprocessed images are unified to a unified geographic coordinate system to obtain the final preprocessed image sequence after geometric positioning; wherein, the initial extrinsic parameters represent the position and attitude data of the aerial camera.

[0010] Furthermore, the process of processing the preprocessed image sequence using structure-of-motion reconstruction (SOR) technology to obtain the camera pose and 3D point cloud of the image includes: Based on machine vision algorithms, feature points are extracted from all images to obtain a number of feature points; A KD-Tree spatial index is constructed based on the initial spatial localization data of the images, and the spatial nearest neighbor images of each image are found according to the spatial index to obtain several image pairs; Based on several feature points, feature matching is performed on several image pairs to obtain initial matching point pairs; The Random Sampling Consensus Algorithm (RANSAC) is used in conjunction with a fundamental matrix or homography matrix model to filter the initial matching point pairs, resulting in purified correct matching point pairs. The fundamental matrix represents the epipolar geometric constraint between two images, which is obtained by applying epipolar constraints to the matching point pairs. The homography matrix model represents the projection transformation under the planar scene, which is obtained by applying a linear transformation to the matching point pairs. Using images as nodes, connect edges are constructed between images with more than a preset number of correctly matched point pairs to obtain an adjacency graph; Based on the adjacency graph, the image pair with the largest information content is selected as the initial seed, and the incremental motion recovery structure process is started to output the camera pose and 3D point cloud of all images; wherein, the information content represents the number of point pairs that are correctly matched between the image pairs.

[0011] Furthermore, the incremental motion recovery structure process is an iterative three-dimensional reconstruction process, including: A1, based on the image pairs of the initial seed, generates an initial 3D point cloud through triangulation to obtain the initial reconstruction model; A2, based on the currently reconstructed 3D point cloud and camera pose, adds new images that match the existing reconstruction model the most and have the best viewpoint conditions, and performs camera pose estimation. A3, perform global bundle adjustment, and perform nonlinear optimization on the camera parameters and coordinates of 3D points in the 3D point cloud of all current images to obtain the optimized camera pose and 3D point cloud; the camera parameters include camera intrinsic parameters and camera extrinsic parameters; A4 uses the new image and the optimized camera pose to triangulate, generating a new 3D point cloud. Steps A2 to A4 are repeated until all new images that meet the conditions are added to the reconstruction model, and the camera pose and 3D point cloud are output.

[0012] Furthermore, the geometric redundancy between the quantized images includes: Based on the camera pose, the optical axis direction vector of the image capturing camera is extracted, and the angle between the cameras is calculated based on the optical axis direction vector to obtain the angle of difference between the images. Based on the three-dimensional point cloud, the three-dimensional Euclidean distance between the two cameras is calculated to obtain the baseline length between the images. The angle of difference and the baseline length are marked as quantitative indicators of geometric redundancy between images, and a preset threshold is set for each indicator. When an indicator exceeds the preset threshold, it is marked as having geometric redundancy.

[0013] Furthermore, the generation of the redundancy comprehensive score includes: Feature vectors of images are extracted using a pre-trained convolutional neural network; Calculate the cosine similarity between the feature vectors of image pairs to obtain a similarity score; Based on the three-dimensional overlap between the three-dimensional point cloud image pairs; A comprehensive redundancy score between image pairs is calculated based on the three-dimensional overlap, the similarity score, the viewpoint difference angle, and the baseline length; the comprehensive redundancy score... The formula for calculation is: ;in, This represents the normalized value. For three-dimensional overlap, For similarity scores, For the angle of difference in viewpoint, Baseline length , , , These are the weighting coefficients for each item.

[0014] Furthermore, the method for calculating the three-dimensional overlap includes: A common-view matrix is ​​established based on the three-dimensional point cloud to obtain the set of three-dimensional points that are jointly observed between any two images; Based on the jointly observed set of three-dimensional points, the three-dimensional overlap between images is calculated.

[0015] Furthermore, the three-dimensional overlap The formula for calculation is: ;in, Let i be the set of three-dimensional points observed in image i. Let j be the set of three-dimensional points observed in image j. Represents the cardinality of a set.

[0016] Furthermore, the step of retaining the optimal image from each cluster according to preset rules includes: Based on the redundancy comprehensive score, a clustering algorithm is used to identify and mark images that exceed a preset score threshold as redundant images, thus obtaining a redundant image cluster. In each redundant image cluster, one optimal image is retained according to preset rules, and all other images are discarded; the preset rules include, but are not limited to, the highest image resolution, the minimum pixel blur, the closest vertical shooting posture, and the best lighting conditions.

[0017] Secondly, this application provides a computer-readable storage medium storing instructions that, when executed by a processor, cause the processor to implement a remote sensing redundancy data identification method based on multi-view geometry theory, the method comprising: The original aerial image sequence is acquired, and the preprocessed image sequence is processed based on the structure-of-motion reconstruction technique to obtain the camera pose and 3D point cloud of the image. Based on the camera pose and the 3D point cloud, the geometric redundancy between images is quantified; the visual features of the images are extracted, and combined with the geometric redundancy, a comprehensive redundancy score is generated. Based on the redundancy comprehensive score, redundant image clusters in the image sequence are identified, and the best image is retained from each cluster according to preset rules to obtain an optimized image set.

[0018] Compared with the prior art, the beneficial effects of this application are: This application employs a systematic approach, deeply integrating multi-view geometry theory into key steps to significantly improve the accuracy, efficiency, and practicality of redundancy processing for low-altitude remote sensing data. First, in the preprocessing stage, by establishing a precise correspondence between imagery and positioning / pose data, and performing radiometric adjustment and coarse geometric positioning, noise caused by illumination differences and initial pose deviations is effectively eliminated. This provides a high-quality data foundation for subsequent motion-based structure reconstruction, avoiding misjudgments caused by data inconsistencies in traditional methods. Second, in motion-based structure reconstruction, based on feature point extraction, spatial index optimization matching, and an incremental reconstruction process, not only is high-precision reconstruction of camera pose and 3D point clouds achieved, but redundancy judgment is also elevated from a two-dimensional representation to a three-dimensional geometric essence. This overcomes the shortcomings of existing technologies in adapting to changes in viewpoint and scene complexity, making redundancy identification more robust.

[0019] Furthermore, when quantifying geometric redundancy, this application introduces the viewpoint difference angle and baseline length as core indicators. By setting dynamic thresholds, it can accurately distinguish between information-complementary images and geometrically redundant images, thus proactively retaining large baseline images crucial for 3D reconstruction during data cleaning, avoiding model quality degradation caused by blind deletion. Simultaneously, when generating a comprehensive redundancy score, it integrates multi-dimensional information such as 3D overlap and visual semantic features, and employs a weighted balancing strategy to ensure the comprehensiveness and adaptability of the scoring system. Finally, a redundancy cluster cleaning mechanism based on clustering algorithms and preset optimization rules achieves automated and intelligent image selection, significantly reducing manual intervention, improving processing efficiency, and ensuring that the optimized image set, while reducing data volume, directly enhances the robustness and accuracy of subsequent 3D modeling or model training. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart illustrating the remote sensing redundancy data identification method based on multi-view geometry theory provided in this application embodiment; Figure 2 A flowchart illustrating another remote sensing redundancy data identification method based on multi-view geometry theory provided in this application embodiment; Figure 3 A flowchart illustrating another remote sensing redundancy data identification method based on multi-view geometry theory provided in this application embodiment; Figure 4 A schematic diagram illustrating the processing of image sequences based on motion restoration structure technology, provided in an embodiment of this application; Figure 5 The original image distribution before removing redundant sample data is provided in the embodiments of this application; Figure 6 The image distribution after removing redundant sample data using a remote sensing redundant data identification method based on multi-view geometry theory, as provided in the embodiments of this application. Detailed Implementation

[0022] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. "And / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0023] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0024] like Figure 1As shown, to address the technical problems in existing technologies such as insufficient perception of the geometric essence of data redundancy, inability to effectively quantify and evaluate view redundancy, and the disconnect between the redundancy detection process and the model training process, this application provides a remote sensing redundancy data identification method based on multi-view geometry theory. This method includes: The original aerial image sequence is acquired and preprocessed to obtain the preprocessed image sequence. The preprocessed image sequence is processed using motion reconstruction technology to obtain the camera pose and 3D point cloud of the image, and the geometric redundancy between the images is quantified. Visual features of the image are extracted and combined with geometric redundancy to generate a comprehensive redundancy score. Based on the redundancy comprehensive score, redundant image clusters in the image sequence are identified, and the best image is retained from each cluster according to preset rules to obtain an optimized image set.

[0025] Based on this, the method of this application breaks through the limitations of traditional two-dimensional feature judgment, improves the accuracy of redundant identification from the perspective of three-dimensional geometry, and realizes automated data cleaning, reducing manual intervention and computational resource consumption. It retains key information and improves the efficiency and quality of subsequent processing. It is applicable to scenarios that rely on low-altitude remote sensing data, such as digital twins, smart cities, real-scene 3D modeling and precision agriculture.

[0026] like Figure 1 As shown in the embodiments of this application, the remote sensing redundancy data identification method based on multi-view geometry theory includes: S1. Acquire the original aerial image sequence and perform preprocessing to obtain the preprocessed image sequence.

[0027] Preprocessing is used to eliminate problems such as format differences, noise interference, geometric deviations, and uneven illumination in the original image, and to standardize data quality.

[0028] In some implementations, preprocessing methods include, but are not limited to: 1. Image format standardization processing: Convert the original aerial photography into a unified uncompressed TIFF format, which may contain multiple formats such as JPG, PNG, and TIFF. This format supports high-resolution image storage and can retain complete pixel information, which is convenient for subsequent feature extraction and data calculation. 2. Noise Removal Processing: For noise generated during aerial photography, such as sensor interference and atmospheric scattering, Gaussian filtering can be used for Gaussian noise, median filtering can be used for salt-and-pepper noise, and bilateral filtering can be used if both noise removal and edge preservation are required. The image clarity is improved by filtering out invalid interference information. 3. Geometric correction processing: using ground control point data, i.e., GCP data, and algorithms such as affine transformation and polynomial correction, to eliminate geometric deviations caused by terrain undulations and camera lens distortion; 4. Lighting equalization processing: To address the uneven brightness of images caused by differences in lighting at different shooting times, gamma correction can be used to adjust the overall brightness, or adaptive histogram equalization can be used to optimize local contrast, so that the brightness and detail of the image tend to be consistent. 5. Image cropping and stitching preprocessing: If the original image has invalid edge areas, it can be cropped to retain the valid scene parts. If there are image breaks, preliminary stitching can be performed to ensure the integrity of scene coverage.

[0029] S2. Process the preprocessed image sequence based on motion recovery structure technology to obtain the camera pose and 3D point cloud of the image, and quantify the geometric redundancy between the images.

[0030] Structure from Motion (SfM) is a technique for recovering camera motion trajectories and the 3D structure of a scene from multiple 2D images, without requiring prior knowledge of camera intrinsic parameters or scene information. Camera pose refers to the position and orientation of the camera when capturing an image. Position information describes the camera's coordinates in 3D space, including X-axis, Y-axis, and Z-axis coordinates, while orientation information describes the camera's shooting angle and direction, including roll, pitch, and yaw angles. Geometric redundancy refers to the degree of overlap in the observation range and geometric information of a scene in 3D space caused by similar viewing angles and shooting positions in different images. It is used to accurately determine the redundancy between images from a 3D geometric perspective.

[0031] In some implementations, the camera pose and 3D point cloud of the image can be obtained through the incremental SfM method based on feature matching; a global SfM method can also be used, which first calculates the relative pose between all images, and then obtains the camera pose and 3D point cloud in a unified coordinate system through global optimization; or it can be combined with Simultaneous Localization and Mapping (SLAM) technology to complete camera pose estimation and 3D point cloud construction while acquiring images in real time.

[0032] In some implementations, geometric redundancy can be quantified by calculating the similarity of observation angles between images. Based on the camera pose, the observation direction vectors of each image are obtained, and the angle between the vectors is calculated. The smaller the angle, the closer the observation angles are, and the higher the geometric redundancy. Alternatively, it can be quantified by calculating the overlap ratio of the corresponding 3D point clouds of the images. The higher the overlap ratio, the more the 3D structure of the scene observed by the two images is repeated, and the higher the geometric redundancy. It can also be combined with the spatial distance between cameras. Images that are closer in distance and have similar observation angles usually have higher geometric redundancy.

[0033] For example, for two aerial images of the same building, the angle between the observation direction vectors of the two cameras is calculated to be 5° through camera pose calculation, and the overlap ratio of the three-dimensional points observed by the two cameras is 85% through three-dimensional point cloud computing. Combining these two indicators, the high geometric redundancy between the two can be quantified, indicating that there is significant repetition at the geometric level.

[0034] S3. Extract the visual features of the image and combine them with geometric redundancy to generate a comprehensive redundancy score.

[0035] Visual features are the unique attributes of an image at the pixel level, including information such as color, texture, edges, and shape. They are used to supplement differences in appearance features not covered at the geometric level, avoiding misjudgments due to similar geometric redundancy but significant differences in visual content. The redundancy comprehensive score is a quantitative result that integrates geometric redundancy information and visual feature information. It is used to provide a unified and comprehensive evaluation standard for the degree of redundancy between images, providing a basis for subsequent redundancy cluster identification.

[0036] In some implementations, traditional visual feature extraction algorithms such as HOG and SURF can be used to extract visual feature vectors from images. The Euclidean distance or cosine similarity between the vectors is calculated as the visual similarity score, and then the geometric redundancy metric is weighted and summed according to preset weights to obtain a comprehensive score. Alternatively, deep learning models such as CNN and ViT can be used to extract high-level semantic features from images, and a multi-dimensional evaluation system can be constructed by combining the geometric redundancy index. Machine learning methods such as logistic regression and random forest can be used to train the scoring model to generate a comprehensive redundancy score. Dynamic weights can also be set for geometric redundancy and visual similarity scores according to the application scenario requirements. For example, in scenes with drastic changes in lighting, the weight of visual features can be increased, while in scenes with stable viewing angles, the weight of geometric redundancy can be increased.

[0037] It should be noted that the choice of visual feature extraction algorithm should be adapted to the characteristics of low-altitude remote sensing imagery. For example, the SURF algorithm can be preferred for terrain images with rich textures, while color feature extraction can be emphasized for scenes with significant color differences. At the same time, the rationality of the overall score needs to be adjusted through sample validation to ensure that the scoring results accurately reflect the actual redundancy between images.

[0038] For example, the HOG feature vectors of two images are extracted and the cosine similarity is calculated to be 0.82. Combined with the quantified geometric redundancy of 0.78, the geometric redundancy weight is set to 0.6 and the visual similarity weight is set to 0.4. The weighted calculation yields a comprehensive redundancy score of 0.80, indicating that the two images have a high degree of redundancy.

[0039] S4. Based on the redundancy comprehensive score, identify redundant image clusters in the image sequence, and retain the best image from each cluster according to preset rules to obtain the optimized image set.

[0040] This step aims to filter out redundant images and remove them, while retaining core images that are of significant value for subsequent 3D reconstruction or model training. This achieves a balance between data simplification and quality improvement, avoiding problems such as wasted computing resources, low training efficiency, and model overfitting caused by redundant data.

[0041] In some implementations, clustering algorithms such as K-Means and DBSCAN can be used to identify redundant image clusters. Images with a redundancy score higher than a preset threshold are grouped into the same cluster, with images having similar scores being more likely to be grouped into the same redundant cluster. Alternatively, graph clustering methods can be used, treating images as nodes and using the redundancy score as the connection weight between nodes. Nodes with weights higher than the threshold form connected components, and each connected component represents a redundant image cluster. The preset rules for retaining the best image can be set according to actual needs. In addition to common criteria such as highest image resolution and minimum pixel blur, they can also include maximum scene coverage, clearest key ground features, and optimal exposure conditions. By quantifying and ranking various image indicators, the image with the best overall performance is selected for retention.

[0042] It should be noted that the threshold setting for the redundancy comprehensive score needs to be combined with the specific application scenario and data characteristics. If the threshold is too high, redundant data may not be fully removed, while if the threshold is too low, useful images may be mistakenly deleted. At the same time, the priority of the preset rules can be adjusted according to the needs of subsequent tasks. For example, in 3D reconstruction tasks, "closest to vertical photographic pose" and "large baseline complementary perspective" can be used as high-priority rules, while in model training tasks, "optimal lighting conditions" and "minimum pixel blur" can be used as high-priority rules.

[0043] Based on the above technical solutions, the remote sensing redundancy data identification method based on multi-view geometry theory provided in this application obtains core data at the three-dimensional geometric level through motion reconstruction structure technology, breaking through the limitations of traditional two-dimensional feature judgment. It can effectively resist interference caused by illumination changes, scale differences, and viewpoint rotation, and improve the accuracy and robustness of redundancy identification. Through a comprehensive scoring mechanism that integrates geometric redundancy and visual features, it achieves multi-dimensional and comprehensive redundancy assessment. Through automated clustering identification and optimal image selection, it reduces manual intervention and improves data processing efficiency. At the same time, it ensures that the optimized image set can both reduce the amount of data and retain key information and viewpoint diversity, providing high-quality data support for subsequent accuracy improvement of 3D reconstruction and efficient training of deep learning models. It is widely applicable to various practical application scenarios that rely on low-altitude remote sensing data processing.

[0044] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 2As shown, the above S1 can be implemented through the following S101, S102 and S103, which are explained in detail below: S101. Establish the correspondence between the image and the POS data of the positioning and attitude determination system.

[0045] The Position and Orientation System is the core system for acquiring camera position and attitude information during aerial photography. By establishing a correspondence between images and the system, each original aerial image can be associated with its position and attitude data at the time of shooting, providing a reliable data association basis for subsequent radiation adjustment and geometric coarse positioning.

[0046] In some implementations, the timestamps of the images can be matched with the timestamps of the POS data to ensure a one-to-one correspondence between the images and the corresponding pose data through time synchronization. Alternatively, the unique number in the image file name can be associated with the number field in the POS data file. If there are inconsistencies in the numbering rules, the mapping relationship in the UAV data acquisition log can be used for correction. The accuracy of the correspondence can also be verified by combining the image shooting order with the POS data recording order.

[0047] For example, in a low-altitude drone aerial photography project, the file name format of the original aerial image is "IMG_20240615_093001_001.jpg", where "20240615_093001" is the shooting timestamp and "001" is the image sequence number. The corresponding POS data file records the latitude and longitude, altitude, roll angle, pitch angle, and yaw angle data corresponding to the same timestamp and sequence number. Through dual verification of timestamp and sequence number, a unique correspondence between each image and POS data is successfully established.

[0048] S102. Perform automatic exposure compensation or histogram equalization on the original aerial image to obtain the first pre-processed image after radiometric adjustment.

[0049] Automatic exposure compensation uses algorithms to adaptively adjust the exposure parameters of images, correcting the uneven brightness of images caused by different shooting times and lighting conditions. Histogram equalization adjusts the grayscale distribution of images to enhance image contrast, making details in dark and bright areas clearer. Both are radiometric adjustment operations, and their core function is to reduce the interference of lighting differences on subsequent feature extraction and improve the consistency of brightness and contrast among all images.

[0050] In some implementations, the brightness histogram of the original image can be analyzed first. If the image is too dark (average gray value below 80, gray value range 0-255) or too bright (average gray value above 200), automatic exposure compensation is preferred. If the image contrast is low (gray value distribution is concentrated in the middle range, variance is less than 50), histogram equalization is preferred. Alternatively, a combination of methods can be used: first, automatic exposure compensation is used to adjust the image brightness to a reasonable range, and then histogram equalization is used to enhance the detail contrast.

[0051] S103. Using POS data, unify the initial extrinsic parameters of all first preprocessed images to a unified geographic coordinate system to obtain the final preprocessed image after geometric positioning.

[0052] The initial extrinsic parameters are the position and attitude data of the aerial camera. The position data includes the three-dimensional spatial coordinates when the camera is shooting, and the attitude data includes roll angle, pitch angle, and yaw angle. Unified geographic coordinates refer to the selection of a unified spatial reference frame, commonly including the National Geodetic Coordinate System 2000 and the WGS-84 coordinate system. This step is used to eliminate the coordinate offset caused by the dispersed shooting positions and different initial pose references of different images, so that all the first preprocessed images are under the same spatial reference system.

[0053] In some implementations, if the POS data provides latitude, longitude, and altitude data, it can be converted into Cartesian coordinates in a unified geographic coordinate system using Gauss-Kruger projection or UTM projection. If the POS data provides relative coordinates, it needs to be calibrated using known reference points (such as ground control points) to convert the relative coordinates into absolute geographic coordinates. Then, combined with the attitude angle data in the POS data, the initial extrinsic parameters of the first preprocessed image are mapped to the unified coordinate system using a coordinate transformation matrix to complete the geometric positioning.

[0054] Based on the above technical solutions, S1 establishes a precise correspondence between images and POS data, providing reliable data association for subsequent processing; it eliminates the image quality inconsistencies caused by illumination differences through radiometric adjustment, improving image consistency; and it unifies all images to the same geographic coordinates through geometric positioning, laying a unified spatial reference foundation. The entire preprocessing workflow effectively eliminates noise and biases in the raw data, providing high-quality, standardized data support for subsequent motion reconstruction techniques. It also significantly improves the accuracy and efficiency of subsequent camera pose calculation and 3D point cloud construction, providing crucial assurance for the accuracy and robustness of the entire remote sensing redundancy data identification method.

[0055] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 3 and Figure 4As shown, the above S2 can be implemented through the following S201, S202, S203, S204, S205, and S206, which are explained in detail below: S201. Based on machine vision algorithms, feature points are extracted from all images to obtain several feature points.

[0056] Feature points are sets of pixels in an image that possess uniqueness, stability, and easy recognition. They typically manifest as corner points, edge endpoints, and areas of dense texture, and are the core foundation for establishing relationships between images and achieving subsequent 3D reconstruction. Machine vision algorithms can automatically select robust feature points from images, ensuring effective matching even under conditions of varying lighting and scale.

[0057] In some implementations, the Scale Invariant Feature Transform (SIFT) algorithm can be selected, which achieves scale-space detection by constructing a Gaussian difference pyramid and can extract scale- and rotation-invariant feature points. Alternatively, the Accelerated Robust Feature Transform (SURF) algorithm can be used, which improves feature extraction speed based on integral images and is suitable for scenarios with high processing efficiency requirements. The ORB algorithm can also be used, which combines FAST corner detection and BRIEF descriptors, balancing real-time performance and robustness, and is suitable for computing platforms with limited resources. For high-precision requirements, the SuperPoint algorithm based on convolutional neural networks can be used to extract feature points with deep semantic information.

[0058] It should be noted that the quality of feature point extraction directly affects the subsequent matching accuracy and 3D reconstruction effect, and a balance must be struck between the number of feature points and computational efficiency—too many feature points will increase the computational burden, while too few may lead to matching failure. At the same time, the choice of feature point extraction algorithm should be adapted to the characteristics of low-altitude remote sensing imagery; for example, terrain imagery can focus on texture features, while urban building imagery can focus on corner features.

[0059] S202. Construct a KD-Tree spatial index based on the initial spatial positioning data of the images, and find the spatial nearest neighbor image of each image according to the spatial index to obtain several image pairs.

[0060] KD-Tree, short for K-Dimensional Tree, is a data structure used for fast retrieval of high-dimensional data. By hierarchically dividing the initial spatial positioning data (three-dimensional positions under unified geographic coordinates), it can significantly shorten the search time for nearest neighbor images. The initial spatial positioning data is the camera's initial position information obtained after preprocessing. The core purpose of building the spatial index is to filter out potentially overlapping image pairs, avoiding full matching of all images and significantly improving processing efficiency.

[0061] In some implementations, a KD-Tree can be constructed based on the three-dimensional coordinates (X,Y,Z) of the image, with a set nearest neighbor number K. The Top-K spatial nearest neighbor images of each image can be found through a proximity search algorithm. Alternatively, flight route planning information can be combined to filter out non-adjacent images along the flight route direction based on the KD-Tree search, further reducing the number of image pairs. For large-scale aerial photography data, a partitioned KD-Tree construction approach can be used to reduce the complexity of a single tree and improve retrieval speed.

[0062] It should be noted that the nearest neighbor number K needs to be set to match the flight path overlap. The higher the lateral overlap, the smaller the K value can be to avoid redundant image pairs; the lower the lateral overlap, the larger the K value needs to be to prevent missing valid overlapping images. At the same time, the accuracy of the initial spatial positioning data will affect the effectiveness of the KD-Tree index. If the positioning error is large, the data needs to be optimized first through interpolation correction or other methods.

[0063] For example, the number of nearest neighbors K is determined based on the lateral overlap of the flight path: when the lateral overlap is 60%–70%, K is set to 5–6; when the overlap is 70%–80%, K is set to 3–5; and when the overlap is above 80%, K is set to 2–3. When constructing a KD-Tree by partitioning, the number of images in each partition can be set to 200–500, balancing retrieval efficiency and the number of partitions.

[0064] S203. Based on several feature points, perform feature matching on several image pairs to obtain initial matching point pairs.

[0065] Feature matching is a process of pairing feature points from the same 3D feature by calculating the similarity of feature point descriptors between different images. The initial matched point pairs contain correctly matched interior points and incorrectly matched exterior points, and serve as the raw data for subsequent refinement processing. Feature descriptors are a quantitative representation of the pixel information surrounding a feature point, and their similarity directly reflects the correspondence between feature points.

[0066] In some implementations, a fast nearest neighbor search package, namely the FLANN matcher, can be used, combined with Euclidean distance to calculate the similarity of floating-point descriptors (such as SIFT and SURF), and a distance threshold can be set to filter matching point pairs. For binary descriptors (such as ORB and BRIEF), Hamming distance can be used to measure similarity, and threshold filtering can be used to achieve fast matching. Alternatively, a brute-force matcher can be used to traverse all feature point pairs to calculate similarity, which is suitable for scenarios with a small number of feature points and ensures the completeness of the matching results.

[0067] It is important to note that setting the matching threshold is crucial—an excessively high threshold will retain a large number of outliers, affecting the accuracy of subsequent processing; an excessively low threshold will discard some correct matches, resulting in insufficient matching rate. Furthermore, the type of feature descriptor must correspond to the feature extraction algorithm to avoid matching failures due to descriptor incompatibility.

[0068] S204. Use the Random Sampling Consensus Algorithm (RANSAC) combined with the fundamental matrix or homography matrix model to filter the initial matching point pairs and obtain the purified correct matching point pairs.

[0069] The Random Sample Consensus (RANSAC) algorithm is a robust parameter estimation algorithm used to select interior points that meet model constraints from an initial set of matched point pairs containing a large number of outliers; these are the correctly matched point pairs. The fundamental matrix represents the polar geometric constraints between two images, describing the correspondence between the projected points of 3D points on the two images. It is suitable for scenes with a high proportion of non-planar elements, such as urban building complexes and mountainous terrain. The homography matrix model represents the projection transformation relationship in planar scenes and is suitable for scenes where the planar element accounts for more than 70%, such as farmland, squares, and airport runways.

[0070] In some implementations, for non-planar scenes, RANSAC combined with the fundamental matrix model is used to filter out outliers through epipolar constraints; for planar scenes, RANSAC combined with the homography matrix model is used to refine matching point pairs through projection transformation constraints; an adaptive selection strategy can also be used, which first calculates the planarity index of the image pair, and if the planarity index is higher than 0.7, the homography matrix is ​​selected, otherwise the fundamental matrix is ​​selected.

[0071] It should be noted that the number of iterations and the interior point threshold in RANSAC directly affect the purification effect—insufficient iterations will prevent finding the optimal model parameters, while too many iterations will increase computation time; an excessively large interior point threshold may retain outliers, while an excessively small threshold will lose valid interior points. Furthermore, the solutions for the fundamental matrix and homography matrix must meet the minimum sample size requirement: the fundamental matrix requires at least 8 pairs of matching points, and the homography matrix requires at least 4 pairs of matching points.

[0072] S205. Using images as nodes, construct connecting edges between images with more than a preset number of correctly matched point pairs to obtain an adjacency graph.

[0073] Among them, the adjacency graph is a graph structure that describes the topological overlap relationship of the image set. Nodes represent individual images, and connecting edges represent that there are a sufficient number of correct matching point pairs between images, that is, there is an effective overlapping area. It is used to intuitively present the correlation strength between images and provide a topological basis for subsequent selection of initial seed image pairs and initiation of incremental reconstruction.

[0074] In some implementations, a preset threshold can be set by counting the number of correctly matched point pairs for each pair of images. When the number of matched point pairs exceeds the threshold, an undirected connection edge is constructed between the corresponding image nodes. Alternatively, a weight can be assigned to the connection edge, with the weight value being the ratio of the number of correctly matched point pairs to the total number of feature points in the two images, which more accurately reflects the degree of overlap. For large-scale image sets, a sparse graph storage method can be used, which only records the node pairs with connection edges, reducing storage overhead.

[0075] It should be noted that the preset threshold for correctly matched point pairs is determined based on the average number of feature points per image, typically set to 20%–30% of the average number of feature points. If an average of 1000 feature points are extracted per image, the threshold is set to 200–300 pairs. When the overlap of flight paths is less than 60%, the threshold can be reduced to 150–200 pairs. The formula for calculating the weight of the connecting edge is: Weight = Number of correctly matched point pairs / (Number of feature points in image A + Number of feature points in image B) × 2, ensuring that the weight value is between 0 and 1.

[0076] S206. Based on the adjacency graph, select the image pair with the most information as the initial seed, start the incremental motion recovery structure process, and output the camera pose and 3D point cloud of all images.

[0077] Among them, the image pair with the most information refers to the image pair with the most correctly matched point pairs in the adjacency graph. Selecting it as the initial seed ensures the stability and accuracy of the initial 3D reconstruction, providing a reliable foundation for adding images in subsequent iterations. The incremental motion reconstruction process is an iterative 3D reconstruction process. Its core is to gradually add new images, optimize camera pose and 3D point clouds, and ultimately achieve globally consistent 3D structure reconstruction of the scene, such as... Figure 4 The image shown is a schematic diagram of image sequence processing based on motion reconstruction technology.

[0078] In some implementations, the incremental motion recovery structure process includes the following steps: First, based on the aforementioned image association map, the image pair with the most information content and geometric stability is selected as the initial reconstruction seed by evaluating the number of matching points and geometric relationships. Then, it enters the incremental main loop: Based on the currently reconstructed 3D point cloud and camera pose, the system efficiently solves the pose estimation problem, selects and adds new images that best match the existing model and have the best viewpoint conditions, and estimates their camera pose. Next, triangulation is performed using the newly added images and existing poses to generate new three-dimensional spatial points, while a triangulation filtering strategy is used to ensure the quality of the new points. After each iteration, a global bundle adjustment is performed to jointly optimize all camera parameters (intrinsic and extrinsic) and 3D point coordinates nonlinearly, thereby effectively controlling error accumulation and ensuring reconstruction accuracy. This iterative process continues until all eligible images are successfully incorporated into the model, ultimately outputting a high-precision camera pose parameters and a sparse 3D point cloud model of the scene.

[0079] For example, assuming that a certain image pair in the adjacency graph has 420 correctly matched point pairs, it is selected as the initial seed; an initial reconstruction model containing 3000 3D points is generated through triangulation; each time a new image is added, it is required that there are ≥150 matching point pairs with the existing model and the viewpoint difference angle is between 15° and 45°. After adding, global bundle adjustment is performed until the root mean square error of reprojection is reduced to 0.4 pixels; finally, 200 images are processed, and the accurate camera pose of all images and a sparse point cloud containing 85,000 3D points are output, providing data support for subsequent geometric redundancy quantification.

[0080] Based on the above technical solution, S2 completed a full motion reconstruction structure processing flow: from feature point extraction to establish local image correlation, to optimizing matching efficiency through KD-Tree indexing, then RANSAC purification to ensure matching accuracy, constructing the global image topology through adjacency graphs, and finally achieving high-precision reconstruction of camera pose and 3D point cloud through an incremental reconstruction process. The entire process deeply integrates multi-view geometry theory, elevating redundancy judgment from a two-dimensional appearance to a three-dimensional geometric essence, effectively overcoming the shortcomings of traditional methods in adapting to changes in viewpoint and scene complexity. Simultaneously, through adaptive parameter adjustment and automated process design, robustness and efficiency are ensured, providing high-quality data for subsequent geometric redundancy quantification and redundancy comprehensive scoring, laying a solid foundation for the accuracy of the entire remote sensing redundancy data identification method.

[0081] In one possible implementation of this application embodiment, the above-mentioned S3 can be specifically implemented by the following S301, S302 and S303, which are described in detail below: S301. Extract the visual feature vector of the image using a pre-trained convolutional neural network.

[0082] Visual feature vectors are high-dimensional quantitative representations of deep semantic information and surface visual features of images, containing key information such as color, texture, shape, and land cover categories. They provide a basis for comparison in subsequent calculations of visual similarity between images. Pre-trained convolutional neural networks refer to network models trained on large general-purpose image datasets (such as ImageNet) or remote sensing-specific datasets (such as NWPU-RESISC45). They possess powerful feature extraction and generalization capabilities and do not require retraining for specific low-altitude remote sensing scenarios.

[0083] In some implementations, ResNet residual networks, such as ResNet50 and ResNet101, can be selected. Their skip-connection structure can effectively alleviate the gradient vanishing problem in deep networks and is suitable for scenarios requiring high-precision features. Visual geometry networks, such as VGG16 and VGG19, can also be used. Their multi-layer convolutional and pooling structures can extract multi-scale visual features and are suitable for low-altitude remote sensing images with rich textures. EfficientNet, an efficient network, can also be selected. It balances model accuracy and computational efficiency through a composite scaling strategy and is suitable for large-scale image processing scenarios. For small-sample low-altitude remote sensing data, a transfer learning strategy can be used. Based on the pre-trained model, the bottom convolutional layers are frozen and the top fully connected layers are fine-tuned to improve the targeting of feature extraction.

[0084] In some implementations, the input image size can be determined based on the model's default configuration. ResNet and VGG series models commonly use 224×224 or 256×256 pixels, while EfficientNet can adapt to 240×240 to 672×672 pixels. Pixel value standardization uses Z-Score standardization (mean μ=0, standard deviation σ=1) or Min-Max standardization, prioritizing the standardization method consistent with that used during pre-trained model training. Model selection depends on scenario requirements: ResNet101 and EfficientNet-B4 are used for high-precision scenarios, ResNet50 and EfficientNet-B0 are used for efficient processing scenarios, and ResNet34 combined with transfer learning is used for small sample scenarios.

[0085] S302. Calculate the similarity score and three-dimensional overlap between image pairs.

[0086] Among them, the similarity score is a quantitative result of the similarity between the visual feature vectors of two images, reflecting the degree of repetition of the images at the visual content and semantic level; the three-dimensional overlap is a quantitative indicator based on the co-view relationship of three-dimensional point clouds, which objectively reflects the degree of repetition of the observation range of two images in three-dimensional space. Together, they provide the core basis for the redundancy comprehensive score at the visual and geometric levels.

[0087] In some implementations, similarity scores can be obtained by calculating the cosine similarity between feature vectors. The cosine similarity value ranges from -1 to 1, with values ​​closer to 1 indicating greater visual similarity. Alternatively, Euclidean distance or Manhattan distance can be used; smaller distances indicate higher similarity, but these values ​​need to be normalized to the [0,1] interval for fusion. For high-dimensional feature vectors, a combination of cosine similarity and Pearson correlation coefficient can be used to improve the robustness of similarity assessment. The calculation of 3D overlap requires first establishing a common-view matrix based on 3D point clouds, recording the set of 3D points jointly observed by any two images, and then using the formula... Calculation, where and Let i and j be the sets of three-dimensional points observed in images i and j, respectively, and |·| represent the cardinality of the set, i.e., the number of elements in the set.

[0088] It should be noted that the key to calculating 3D overlap is the accuracy of the common-view matrix. The precision and integrity of the 3D point cloud must be ensured. If the 3D point cloud contains a large number of noisy points, noise must be removed first through statistical filtering, radius filtering, or other methods; otherwise, the overlap calculation results will be affected. Furthermore, the 3D overlap value ranges from [0,1], where 0 represents no overlap and 1 represents complete overlap. This index effectively penetrates visual appearance differences and reflects the redundant nature at the geometric level.

[0089] S303. Generate a comprehensive redundancy score based on the three-dimensional overlap, similarity score, viewing angle difference angle, and baseline length.

[0090] Among them, the redundancy comprehensive score is a unified quantitative indicator that integrates geometric redundancy information (three-dimensional overlap, viewing angle difference angle, baseline length) and visual semantic information (similarity score). It is used to provide a comprehensive and objective evaluation standard for the redundancy of image pairs. The higher the score, the greater the possibility of redundancy in the image pair, providing a clear basis for subsequent identification of redundant image clusters.

[0091] In some implementations, the formula for calculating the overall redundancy score is as follows: ,in This represents the normalized value. For three-dimensional overlap, For similarity scores, The angle of difference in view is obtained by calculating the angle between the two cameras; The baseline length is obtained by calculating the three-dimensional Euclidean distance between the two cameras; These are adjustable weighting coefficients, and their sum is 1, used to balance the influence weights of each indicator.

[0092] In some implementations, the weighting coefficient The importance of each indicator can be determined through the analytic hierarchy process (AHP), by inviting domain experts to score the importance of each indicator, constructing a judgment matrix, and calculating the weights. Alternatively, a grid search method can be used to optimize the weight values ​​based on the redundancy recognition accuracy of the validation set. Furthermore, the weights can be dynamically adjusted according to the application scenario; for example, in a 3D reconstruction scenario, increasing the weights... (Baseline length weighting) prioritizes preserving large baseline images; in model training scenarios, this is increased. (Similarity score weights) to reduce visual redundancy.

[0093] Based on the above technical solutions, S301 extracts high-dimensional visual feature vectors through a pre-trained convolutional neural network, ensuring the depth and robustness of visual information capture; S302 quantifies from two core dimensions, visual similarity and three-dimensional geometric overlap, making up for the one-sidedness of single-dimensional evaluation, taking into account both the differences in image appearance and the essence of geometric redundancy; S303 integrates multi-dimensional indicators through a weighted fusion strategy to generate a unified comprehensive redundancy score, realizing the organic combination of geometric information and visual semantic information, providing a reliable basis for the accurate identification of subsequent redundant image clusters.

[0094] In one possible implementation of this application embodiment, the above-mentioned S4 specifically includes the following S401 to S403: S401. Determine the preset scoring threshold for the redundancy comprehensive score.

[0095] The preset scoring threshold is a critical value that distinguishes whether an image pair is redundant. It is used to ensure that only images with a redundancy level that meet the set requirements can be included in the redundant image cluster, thus avoiding misjudging non-redundant images or missing truly redundant images.

[0096] In some implementations, the threshold can be determined using statistical distribution methods: Calculate the statistical distribution of the redundancy comprehensive score for all images, and use the 90th or 95th percentile of the score as a preset threshold, which is applicable to most general low-altitude remote sensing scenarios. Alternatively, a validation set calibration method can be used: select a validation set containing known redundant and non-redundant images, iterate through different thresholds (such as 0.4 to 0.8), and use the value that makes the redundancy recognition accuracy the highest as the final threshold. This method is suitable for scenarios with high accuracy requirements. Another approach is scene adaptation: adjust the threshold according to the shooting scene of the image (such as urban buildings, farmland, and mountains). For urban building scenes, which have high texture repetition, the threshold can be set to 0.65 to 0.75. For farmland and mountain scenes, which have complex terrain, the threshold can be reduced to 0.5 to 0.6.

[0097] S402. Based on a preset scoring threshold and clustering algorithm, identify redundant image clusters in the image sequence.

[0098] Redundant image clusters refer to image sets whose overall redundancy score exceeds a preset threshold and whose redundancy is high among themselves. The core characteristic is that the images within the cluster have high repetition of observations of 3D ground features and contribute minimal stereoscopic information. The clustering algorithm automatically groups images that meet the redundancy criteria, enabling batch identification of redundant images.

[0099] In some implementations, a density-based noisy spatial clustering algorithm, DBSCAN, can be used. This algorithm does not require a preset number of clusters, can automatically identify clusters of arbitrary shapes, and can filter non-redundant images, making it suitable for scenarios with irregular distribution of redundant images. Alternatively, the K-Means clustering algorithm can be used. First, the number of clusters K is preset according to the size of the image set and the redundancy distribution (usually K is 1 / 5 to 1 / 3 of the total number of images). Then, images with scores exceeding a threshold are assigned to the nearest cluster center, making it suitable for scenarios with relatively uniform distribution of redundant images. Another option is a hierarchical clustering algorithm. This algorithm constructs a cluster tree by calculating the similarity of redundancy between images, and cuts the tree structure according to a preset threshold to obtain redundant image clusters, making it suitable for scenarios where the hierarchical relationship between clusters needs to be clearly defined.

[0100] S403. According to the preset rules, retain one optimal image from each redundant image cluster and remove the remaining redundant images to obtain the optimized image set.

[0101] The preset rules are the specific criteria for selecting the best images. The core objective is to retain the images that contribute the most to subsequent 3D reconstruction or model training from the redundant cluster, ensuring that the optimized image set is both streamlined in terms of data volume and fully preserves key scene information and stereoscopic observation value. Optimal images typically possess characteristics such as high resolution, low blurriness, excellent viewing angle, and good lighting conditions.

[0102] In some implementations, a single rule priority strategy can be adopted: Prioritize the selection of images with the highest resolution. If the resolutions are the same, select the image with the lowest pixel blur. This is suitable for model training scenarios that require high image clarity. The lowest pixel blur indicates that the diffusion of a single pixel in the remote sensing image is the lowest, with sharp edges and accurate detail reproduction. It is free from blurring caused by factors such as drone shaking, improper camera exposure, or focus deviation. It can preserve key visual information such as texture and edges of ground objects to the greatest extent, providing high-definition basic data for model training. A multi-dimensional weighted scoring strategy can also be adopted: image resolution, pixel blur, vertical photography pose deviation, and illumination uniformity are quantified into indicators of 0 to 1 points, and weights are assigned (resolution 0.3, blur 0.3, pose deviation 0.2, illumination 0.2). The weighted total score of each image is calculated, and the image with the highest total score is retained. This is suitable for scenarios such as 3D reconstruction that require multi-dimensional features. Scene customization rules can also be adopted: urban building scenes prioritize images that are closest to the vertical shooting posture, farmland scenes prioritize images with the best lighting uniformity, and mountain scenes prioritize images whose perspective covers key terrain features.

[0103] Based on the above technical solutions, S4, through threshold determination, cluster grouping, and optimal image selection, deeply meets the needs of subsequent 3D reconstruction or model training. It not only solves the problem of redundancy detection being disconnected from the main task in existing methods, but also realizes intelligent and accurate redundancy cleaning, laying a high-quality data foundation for the efficient implementation of subsequent processes, and further enhancing the practicality and engineering value of the entire remote sensing redundancy data identification method.

[0104] Based on the above technical solution, the processing effect of the redundant data identification method of this application is as follows: Figure 5 and Figure 6 As shown, where, Figure 5 Image distribution before removing redundant sample data, Figure 6 This describes the image distribution after removing redundant sample data. Therefore, the low-altitude remote sensing sample redundancy data identification method based on multi-view geometry theory provided by this invention offers significant and multifaceted improvements compared to existing technologies, specifically in the following aspects: First, it achieves precise identification from appearance to essence, significantly improving the accuracy of redundancy removal. This invention transcends the limitations of traditional two-dimensional image similarity comparison, directly restoring the precise three-dimensional pose and scene structure of images through Structure for Motion Restoration (SfM), thus determining data redundancy from a three-dimensional geometric perspective. This allows the system to effectively penetrate apparent differences such as lighting variations and viewpoint rotation, accurately identifying images that have been repeatedly observed for the same three-dimensional feature. Simultaneously, it avoids mistakenly deleting key images with similar textures but different spatial locations, fundamentally solving the problem of high misjudgment rates in traditional methods, resulting in more reliable cleaning results.

[0105] Secondly, this invention enables a quantitative assessment of the value of stereoscopic information, ensuring and optimizing the quality of 3D reconstruction while compressing data volume. The invention innovatively introduces viewpoint difference and baseline length as core criteria, effectively distinguishing between essential images with complementary information and truly redundant images with geometric repetition. This method proactively retains large baselines and unique viewpoint images that contribute more to stereoscopic measurement and model accuracy during the cleaning process, avoiding indiscriminate deletion. The dataset processed by this invention not only significantly reduces data volume but also enhances the accuracy and completeness of the constructed 3D model, achieving a leap from quantity reduction to quality improvement.

[0106] Third, it achieves seamless integration with the 3D reconstruction workflow, significantly improving overall processing efficiency. This invention is based on the precise geometric information produced by the SfM workflow, and its redundancy judgment criteria are highly consistent with the inherent requirements of the 3D reconstruction engine. The cleaned, high-quality image subset has stronger connectivity between images, effectively avoiding SfM reconstruction failures or model fragmentation caused by data cleaning. Simultaneously, due to the sharp reduction in the amount of data input to SfM and subsequent dense matching, the computation time and memory usage of the entire 3D reconstruction workflow can be reduced by more than 50%, greatly saving computational resources and time costs.

[0107] Fourth, it possesses excellent automation and practicality, making it suitable for large-scale engineering applications. By constructing spatial indexes for initial matching and screening, automatically quantizing based on common-view relationships, and determining dynamic thresholds, the entire process minimizes manual intervention and achieves automated operation. This method can efficiently process large-scale low-altitude datasets consisting of thousands or even tens of thousands of images, providing crucial technical support for the automated processing of massive amounts of data in large-scale projects such as real-scene 3D and digital twins.

[0108] In summary, this invention successfully applies multi-view geometry theory to the engineering challenge of redundancy processing of low-altitude remote sensing data, achieving a synergistic improvement in accuracy, quality, efficiency, and practicality, and providing an effective solution for the intelligent and refined processing of low-altitude remote sensing data.

[0109] In implementation, each step of the method provided in this embodiment can be completed by integrated logic circuits in the processor or by instructions in software form. The steps of the method disclosed in the embodiments of this application can be directly manifested as being executed by a hardware processor, or being executed by a combination of hardware and software modules in the processor.

[0110] This application also provides a computer-readable storage medium including instructions that, when run on a computer, cause the computer to perform any of the methods described above.

[0111] This application also provides a computer program product containing instructions that, when run on a computer, cause the computer to perform any of the methods described above.

[0112] This application also provides a chip including a processor and an interface circuit. The interface circuit is coupled to the processor. The processor is used to run computer programs or instructions to implement the above-described method. The interface circuit is used to communicate with other modules outside the chip.

[0113] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0114] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and modifications of this application fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and modifications.

Claims

1. A method for identifying redundant remote sensing data based on multi-view geometry theory, characterized in that, include: The original aerial image sequence is acquired and preprocessed to obtain the preprocessed image sequence. The preprocessed image sequence is processed using motion recovery structure technology to obtain the camera pose and 3D point cloud of the image, and the geometric redundancy between the images is quantified. Visual features of the image are extracted and combined with the geometric redundancy to generate a comprehensive redundancy score. Based on the redundancy comprehensive score, redundant image clusters in the image sequence are identified, and the optimal image is retained from each cluster according to preset rules to obtain an optimized image set; wherein... The process of processing the preprocessed image sequence using motion reconstruction technology to obtain the camera pose and 3D point cloud of the image includes: Based on machine vision algorithms, feature points are extracted from all images to obtain a number of feature points; A KD-Tree spatial index is constructed based on the initial spatial localization data of the images, and the spatial nearest neighbor images of each image are found according to the spatial index to obtain several image pairs; Based on several feature points, feature matching is performed on several image pairs to obtain initial matching point pairs; The Random Sampling Consensus Algorithm (RANSAC) is used in conjunction with a fundamental matrix or homography matrix model to filter the initial matching point pairs, resulting in purified correct matching point pairs. The fundamental matrix represents the epipolar geometric constraint between two images, which is obtained by applying epipolar constraints to the matching point pairs. The homography matrix model represents the projection transformation under the planar scene, which is obtained by applying a linear transformation to the matching point pairs. Using images as nodes, connect edges are constructed between images with more than a preset number of correctly matched point pairs to obtain an adjacency graph; Based on the adjacency graph, the image pair with the largest information content is selected as the initial seed, and the incremental motion recovery structure process is started to output the camera pose and 3D point cloud of all images; where the information content represents the number of point pairs that are correctly matched between the image pairs. The geometric redundancy between the quantized images includes: Based on the camera pose, the optical axis direction vector of the image capturing camera is extracted, and the angle between the cameras is calculated based on the optical axis direction vector to obtain the angle of difference between the images. Based on the three-dimensional point cloud, the three-dimensional Euclidean distance between the two cameras is calculated to obtain the baseline length between the images. The angle of difference and the baseline length are marked as quantitative indicators of geometric redundancy between images, and a preset threshold is set for each indicator. When an indicator exceeds the preset threshold, it is marked as having geometric redundancy.

2. The remote sensing redundancy data identification method based on multi-view geometry theory according to claim 1, characterized in that, The incremental motion recovery structure process is an iterative three-dimensional reconstruction process, including: A1, based on the image pairs of the initial seed, generates an initial 3D point cloud through triangulation to obtain the initial reconstruction model; A2, based on the currently reconstructed 3D point cloud and camera pose, adds new images that match the existing reconstruction model the most and have the best viewpoint conditions, and performs camera pose estimation. A3, perform global bundle adjustment, and perform nonlinear optimization on the camera parameters and coordinates of 3D points in the 3D point cloud of all current images to obtain the optimized camera pose and 3D point cloud; the camera parameters include camera intrinsic parameters and camera extrinsic parameters; A4 uses the new image and the optimized camera pose to triangulate, generating a new 3D point cloud. Steps A2 to A4 are repeated until all new images that meet the conditions are added to the reconstruction model, and the camera pose and 3D point cloud are output.

3. The remote sensing redundancy data identification method based on multi-view geometry theory according to claim 1, characterized in that, The generated redundancy comprehensive score includes: Feature vectors of images are extracted using a pre-trained convolutional neural network; Calculate the cosine similarity between the feature vectors of image pairs to obtain a similarity score; Based on the three-dimensional overlap between the three-dimensional point cloud image pairs; A comprehensive redundancy score between image pairs is calculated based on the three-dimensional overlap, the similarity score, the viewpoint difference angle, and the baseline length; the comprehensive redundancy score... The formula for calculation is: ;in, This represents the normalized value. For three-dimensional overlap, For similarity scores, For the angle of difference in viewpoint, Baseline length , , , These are the weighting coefficients for each item.

4. The remote sensing redundancy data identification method based on multi-view geometry theory according to claim 3, characterized in that, The method for calculating the three-dimensional overlap includes: A common-view matrix is ​​established based on the three-dimensional point cloud to obtain the set of three-dimensional points that are jointly observed between any two images; Based on the jointly observed set of three-dimensional points, the three-dimensional overlap between images is calculated.

5. The remote sensing redundancy data identification method based on multi-view geometry theory according to claim 4, characterized in that, The three-dimensional overlap The formula for calculation is: ;in, Let i be the set of three-dimensional points observed in image i. Let j be the set of three-dimensional points observed in image j. Represents the cardinality of a set.

6. The remote sensing redundancy data identification method based on multi-view geometry theory according to claim 1, characterized in that, The step of retaining the optimal image from each cluster according to preset rules includes: Based on the redundancy comprehensive score, a clustering algorithm is used to identify and mark images that exceed a preset score threshold as redundant images, thus obtaining a redundant image cluster. In each redundant image cluster, one optimal image is retained according to preset rules, and all other images are discarded; the preset rules include the highest image resolution, the minimum pixel blur, the closest vertical shooting pose, and the best lighting conditions.

7. The remote sensing redundancy data identification method based on multi-view geometry theory according to claim 1, characterized in that, The preprocessing steps for the original aerial image sequence include: Establish the correspondence between images and POS data from the positioning and attitude determination system; Automatic exposure compensation or histogram equalization is performed on the original aerial image to obtain the first pre-processed image after radiometric adjustment. Using POS data, the initial extrinsic parameters of all first preprocessed images are unified to a unified geographic coordinate system to obtain the final preprocessed image after geometric positioning; wherein, the initial extrinsic parameters represent the position and attitude data of the aerial camera.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by a processor, cause the processor to implement a remote sensing redundancy data identification method based on multi-view geometry theory, the method comprising: The original aerial image sequence is acquired and preprocessed to obtain the preprocessed image sequence. The preprocessed image sequence is processed using motion recovery structure technology to obtain the camera pose and 3D point cloud of the image. Based on the camera pose and the 3D point cloud, the geometric redundancy between images is quantified; the visual features of the images are extracted, and combined with the geometric redundancy, a comprehensive redundancy score is generated. Based on the redundancy comprehensive score, redundant image clusters in the image sequence are identified, and the optimal image is retained from each cluster according to preset rules to obtain an optimized image set; wherein... The process of processing the preprocessed image sequence using motion reconstruction technology to obtain the camera pose and 3D point cloud of the image includes: Based on machine vision algorithms, feature points are extracted from all images to obtain a number of feature points; A KD-Tree spatial index is constructed based on the initial spatial localization data of the images, and the spatial nearest neighbor images of each image are found according to the spatial index to obtain several image pairs; Based on several feature points, feature matching is performed on several image pairs to obtain initial matching point pairs; The Random Sampling Consensus Algorithm (RANSAC) is used in conjunction with a fundamental matrix or homography matrix model to filter the initial matching point pairs, resulting in purified correct matching point pairs. The fundamental matrix represents the epipolar geometric constraint between two images, which is obtained by applying epipolar constraints to the matching point pairs. The homography matrix model represents the projection transformation under the planar scene, which is obtained by applying a linear transformation to the matching point pairs. Using images as nodes, connect edges are constructed between images with more than a preset number of correctly matched point pairs to obtain an adjacency graph; Based on the adjacency graph, the image pair with the largest information content is selected as the initial seed, and the incremental motion recovery structure process is started to output the camera pose and 3D point cloud of all images; where the information content represents the number of point pairs that are correctly matched between the image pairs. The geometric redundancy between the quantized images includes: Based on the camera pose, the optical axis direction vector of the image capturing camera is extracted, and the angle between the cameras is calculated based on the optical axis direction vector to obtain the angle of difference between the images. Based on the three-dimensional point cloud, the three-dimensional Euclidean distance between the two cameras is calculated to obtain the baseline length between the images. The angle of difference and the baseline length are marked as quantitative indicators of geometric redundancy between images, and a preset threshold is set for each indicator. When an indicator exceeds the preset threshold, it is marked as having geometric redundancy.

Images