Lightweight remote sensing image geometric precise correction method and system based on B / S architecture

Abstract

The present application relates to a lightweight remote sensing image geometric precise rectification method and system based on B / S architecture, and belongs to the technical field of remote sensing data processing, which solves the problems of complex deployment and low efficiency existing in the prior art. The method first constructs a GCP database; based on the database, a homonym sample library is constructed and a feature extraction network is trained; the to-be-rectified image is subjected to resolution recognition, candidate GCP screening and GCP candidate search window division; based on the trained feature extraction network and similarity, homonym matching is carried out, the homonym pixel coordinates are obtained and visualized display is realized through the browser front end; the homonym sample library is updated according to the artificial checking operation instruction; the to-be-rectified image is subjected to geometric precise rectification and resampling, and the rectification result is returned to the browser front end for visualized display. The present application can run in a general computer and browser environment, realize lightweight, low-cost and high-efficiency remote sensing image geometric precise rectification, and is simple to deploy and flexible to upgrade.

Description

Technical Field

[0001] This invention relates to the field of remote sensing data processing technology, and in particular to a lightweight method and system for geometric fine correction of remote sensing images based on a B / S architecture. Background Technology

[0002] Geometric correction of remote sensing images is a crucial step in remote sensing data processing. It aims to eliminate or reduce distortions caused by sensor attitude, terrain undulation, and imaging geometry using ground control points and mathematical models, enabling the image to align with the true geographic coordinate system. Geometrically corrected orthophotos are widely used in mapping, land surveys, urban planning, environmental monitoring, and Geographic Information System (GIS) overlay analysis, requiring extremely high spatial positioning accuracy.

[0003] Currently, the mainstream methods for geometric correction of remote sensing images in the industry mainly rely on large-scale professional software such as PCI Geomatica, ArcGIS, and ENVI. While these software programs are comprehensive in function, they still have significant shortcomings:

[0004] High cost: Software licenses are expensive, and additional module licenses are often required, increasing the cost of use;

[0005] High maintenance costs: The deployment environment is complex, highly dependent on hardware and system environment, and upgrades and maintenance require professional personnel support, resulting in high long-term operating costs;

[0006] Low collaboration efficiency: Most software is desktop or client / server architecture, which makes it difficult to achieve efficient collaborative processing between different departments or teams and cannot meet the needs of large-scale, multi-user parallel operations.

[0007] Insufficient update flexibility: Traditional software lacks an open sample library management and reuse mechanism, making it difficult to quickly adapt and optimize for new technology iterations such as automatic homonym identification.

[0008] In recent years, although deep learning and automatic matching methods have been gradually applied to the identification of corresponding points, related research has mostly remained at the algorithm level and has not yet formed a lightweight solution that can be engineered and promoted. It still relies on high-performance computing environments or large professional platforms, resulting in problems such as complex deployment, high cost and low efficiency in practical applications.

[0009] Therefore, there is an urgent need for a geometric correction method based on a lightweight architecture that can run on general-purpose computers and browsers, avoid dependence on large commercial software, reduce deployment and maintenance costs, improve collaboration capabilities and processing efficiency, and support closed-loop updates of the sample library to continuously improve the accuracy of corresponding point identification and geometric correction. Summary of the Invention

[0010] The purpose of this invention is to overcome the problems of complex deployment, high cost, low efficiency, and poor upgrade flexibility in existing remote sensing image geometric correction technologies, and to provide a lightweight remote sensing image geometric precision correction method and system based on a B / S architecture. This invention, through the combination of browser front-end visualization and server back-end computation, can achieve automatic matching and geometric precision correction of image corresponding points, avoiding dependence on large commercial software, and featuring low cost, easy deployment, and scalability.

[0011] To achieve the above objectives, the present invention adopts the following technical solution:

[0012] A lightweight geometrical correction method for remote sensing images based on a B / S architecture is implemented. The method operates within a B / S architecture consisting of a browser front-end and a server back-end. The server back-end performs the following steps:

[0013] Step 1: Construct a GCP database, which stores the geographic coordinates of several GCPs, and the GCP template images and their spatial resolutions at three different spatial resolutions of 0.5 m / pixel, 0.75 m / pixel and 1.0 m / pixel for each GCP. The size of the GCP template image is 512×512 pixels.

[0014] Step 2: Construct a corresponding point sample library based on the GCP database, and use the corresponding point sample library to train and optimize the parameters of the feature extraction network;

[0015] Step 3: Acquire the image to be corrected and preprocess the image to be corrected, including identifying the resolution, filtering candidate GCPs, and defining a GCP candidate search window for each candidate GCP in the image to be corrected;

[0016] Step 4: Using the trained feature extraction network, extract the feature vectors of the GCP template image and each local search sub-image with a size of 512×512 pixels in the GCP candidate search window for each candidate GCP. Calculate the similarity between the feature vectors of the GCP template image and each local search sub-image. After the local search sub-image traverses the GCP candidate search window, calculate the pixel coordinates of the corresponding point of the candidate GCP based on the position of the optimal local search sub-image corresponding to the maximum similarity. Send the pixel coordinates of the corresponding point and the corresponding GCP candidate search window and GCP template image to the browser front-end for visualization.

[0017] Step 5: Receive manual review instructions from the browser front-end, and update the same-name point sample library according to the manual review instructions;

[0018] Step 6: Based on the pixel coordinates of the corresponding points confirmed by the manual review operation command, perform geometric fine correction and resampling on the image to be corrected, and return the correction results to the browser front end for visualization display.

[0019] Accordingly, the present invention also provides a lightweight remote sensing image geometric fine correction system based on a B / S architecture for implementing the method described above. The system adopts a B / S architecture consisting of a browser front-end and a server back-end, and includes:

[0020] Deployed on the backend of the server:

[0021] The GCP management module is used to build a GCP database. The GCP database stores the geographic coordinates of several GCPs, and the GCP template images and their spatial resolutions under three different spatial resolution specifications of 0.5 meters / pixel, 0.75 meters / pixel and 1.0 meters / pixel for each GCP. The size of the GCP template image is 512×512 pixels.

[0022] The sample library construction and training module is used to construct a sample library of corresponding points based on the GCP database, and to train and optimize the parameters of the feature extraction network using the sample library of corresponding points.

[0023] The preprocessing module is used to acquire the image to be corrected and to preprocess the image to be corrected, including identifying the resolution, filtering candidate GCPs, and defining a GCP candidate search window for each candidate GCP in the image to be corrected.

[0024] The corresponding point automatic matching module is used to extract the feature vectors of the GCP template image and each local search sub-image with a size of 512×512 pixels in the GCP candidate search window for each candidate GCP using a trained feature extraction network, and calculate the similarity between the feature vectors of the GCP template image and each local search sub-image. After the local search sub-image traverses the GCP candidate search window, the pixel coordinates of the corresponding point of the candidate GCP are calculated based on the position of the optimal local search sub-image corresponding to the maximum similarity.

[0025] The sample library update module is used to receive manual review operation instructions and update the sample library of the same name points according to the manual review operation instructions.

[0026] The geometric correction module is used to perform geometric fine correction and resampling on the image to be corrected based on the pixel coordinates of the corresponding points confirmed by the manual review operation instructions, and return the correction results to the browser front end for visualization display.

[0027] And, running in the browser front end:

[0028] The human-computer interaction and visualization module is used to visualize the pixel coordinates of the corresponding points output by the automatic matching module and their corresponding GCP candidate search windows and GCP template diagrams through a visualization interface, and to receive interactive operations triggered by the user based on the visualization interface, and generate and output the corresponding manual review operation instructions to the sample library update module.

[0029] This invention achieves low-cost, lightweight, and iterative geometrical correction of remote sensing images, combining engineering practicality and scalability. It effectively solves the problems of high cost, difficult maintenance, and low efficiency in existing technologies. Compared with existing technologies, this invention has the following beneficial effects:

[0030] (1) Lightweight and low cost: The present invention adopts a B / S architecture, which can complete the geometric correction of remote sensing images in a general computer and browser environment without relying on large paid software such as PCI and ArcGIS, which significantly reduces the complexity of system deployment and reduces the cost and threshold of use and maintenance.

[0031] (2) High efficiency and high precision: Based on the trained feature extraction network, the automatic matching of corresponding points is realized, which greatly improves the efficiency of corresponding point matching. At the same time, combined with the browser front-end visualization and manual review, the accuracy of the matching results is ensured.

[0032] (3) Visualization and one-click operation: The correction results are displayed through the browser front end and manual review is supported. Users can intuitively compare images and correct errors in the HTML page to achieve convenient human-computer collaboration.

[0033] (4) Sample library closed-loop mechanism: During the manual review process, incorrect matches are automatically recorded as negative samples, and precision deviation points are recorded as difficult positive samples. These sample data are fed back into the sample library of the same name to provide support for subsequent model retraining and recognition accuracy improvement. Attached Figure Description

[0034] Figure 1 This is a flowchart of the lightweight remote sensing image geometric fine correction method based on B / S architecture according to an embodiment of the present invention;

[0035] Figure 2 A schematic diagram of the GCP candidate search window defined in the image to be corrected;

[0036] Figure 3 This is a schematic diagram showing the correspondence between the pixel coordinates of the corresponding point of the candidate GCP and the geographic coordinates of the candidate GCP.

[0037] Figure 4 This is an architecture diagram of a lightweight remote sensing image geometric correction system based on a B / S architecture, as described in another embodiment of the present invention. Detailed Implementation

[0038] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings and preferred embodiments.

[0039] like Figure 1 As shown, one embodiment of the present invention provides a lightweight remote sensing image geometric fine correction method based on a B / S architecture, the method comprising the following steps:

[0040] Step 1: Create and store Ground Control Points (GCPs) in the database to build the GCP database.

[0041] In this step, high-resolution imagery is first acquired through methods such as UAV aerial surveying. Then, high-precision mapping techniques, including Real-Time Kinematic (RTK) positioning, are used to conduct on-site measurements of the selected high-resolution imagery-covered geographical features (GCPs), thereby obtaining the centimeter-level geographic coordinates of the GCPs. The selection criteria for GCPs are that the geometric features of the ground features are obvious, stable, and not easily changed, such as road intersections, building corners, or artificially placed markers. RTK is a high-precision positioning technology based on Global Navigation Satellite System (GNSS) systems such as GPS, BeiDou, GLONASS, and Galileo.

[0042] After acquiring the original high-resolution imagery, it is uniformly downsampled to a spatial resolution of 0.1 meters per pixel. A 512×512 pixel original reference sub-image of each selected GCP is then cropped, centered on the original image. If the original high-resolution image is smaller than this size, it is padded with black or white. The cropped original reference sub-images of the GCPs are stored in both PNG and TIFF formats. The PNG format is used as thumbnails for quick browsing and manual verification, while the TIFF format is used as the main image for subsequent feature extraction and automatic matching. Both types of images maintain a standard size of 512×512 pixels, and their corresponding spatial resolutions are recorded in the database.

[0043] After creating the 0.1-meter resolution GCP raw reference sub-image, it is further transformed to generate multi-scale versions of the GCP template image. Specifically, the created GCP raw reference sub-image is downsampled to three spatial resolutions: 0.5 meters / pixel, 0.75 meters / pixel, and 1.0 meters / pixel, resulting in three different spatial resolution GCP template images. These will be used to construct a GCP database with the same resolution as the target satellite or aerial image.

[0044] To ensure data integrity and traceability, each GCP template diagram and its related information (see Table 1) are stored uniformly in the GCP database, and the fields of the GCP database are shown in Table 1.

[0045] Table 1 Field Definitions of the GCP Database

[0046]

[0047] Step 2: Construct a sample library of corresponding points based on the GCP database, and use the sample library of corresponding points to train and optimize the parameters of the feature extraction network.

[0048] Step 2.1: Construct a sample library of points with the same name.

[0049] Select multiple remote sensing images containing at least one GCP from step 1, with spatial resolutions of 0.5 m / pixel, 0.75 m / pixel, and 1.0 m / pixel, respectively, as target images. On each target image, crop an image with the same size as the GCP template image (512×512 pixels) centered on the geographical location of the GCP. This yields local area images corresponding to each GCP template image at different spatial resolutions (0.5 m / pixel, 0.75 m / pixel, and 1.0 m / pixel). This allows for the generation of one-to-one image pairs at different spatial resolutions, namely, the GCP template image (at a specific spatial resolution) and the local area image (at the same spatial resolution as the GCP template image).

[0050] Subsequently, at each spatial resolution, an image pair consisting of the GCP template image and its corresponding local region image is used as a corresponding point matching sample. All corresponding point matching samples are stored in a corresponding point sample library, and each sample in this library contains the following information:

[0051] (1) GCP template diagram and its resolution specifications (0.5 m / pixel, 0.75 m / pixel or 1.0 m / pixel);

[0052] (2) Local area image corresponding to the GCP template image;

[0053] (3) The relationship between the GCP template map and the corresponding local area image at the pixel level, that is, the mapping between the pixel coordinates of the local area image and the real geographic coordinates corresponding to the GCP.

[0054] (4) Metadata information, including sample number, resolution level, acquisition time, image source, etc.

[0055] This step establishes a sample library of corresponding points covering multiple spatial resolutions. This sample library not only ensures the consistency of GCP template images and local area images in terms of spatial scale, but also provides standardized and structured sample data support for the subsequent training of feature extraction networks used for automated identification of corresponding points. This significantly improves the adaptability and recognition accuracy of automated identification of corresponding points across images with different spatial resolutions, and enhances the efficiency of geometric correction of remote sensing images.

[0056] Step 2.2: Model training.

[0057] In the automatic identification of corresponding points, this invention employs a feature matching method based on deep learning, which specifically includes the following steps:

[0058] Step 2.2.1: First, input the corresponding point matching sample into the feature extraction network (such as convolutional neural network CNN or Transformer). The feature extraction network extracts the high-dimensional feature vectors of the GCP template map and the local region image from the corresponding point matching sample.

[0059] Step 2.2.2: Calculate the similarity between two feature vectors using a similarity metric function; the formula for calculating the similarity metric function in this step is as follows:

[0060] ;

[0061] in:

[0062] The input is a GCP template image (with a resolution of 0.5 m / pixel, 0.75 m / pixel, or 1.0 m / pixel).

[0063] This is a local area image corresponding to the input GCP template image;

[0064] This is a feature extraction network used to extract the input GCP template image. Or local area image Encode as a feature vector;

[0065] GCP template diagram eigenvectors;

[0066] For local area images eigenvectors;

[0067] This refers to the dot product operation of vectors.

[0068] The L2 norm of the eigenvectors is used for normalization;

[0069] GCP template diagram With local area images The similarity value between them ranges from 1 to 10. The larger the value, the higher the GCP template diagram. With local area images The higher the degree of matching.

[0070] Step 2.2.3: Iterate with the goal of optimizing the parameters of the feature extraction network to maximize the similarity between images that are matched from the same point of origin, and minimize the similarity between images that are matched from different points of origin.

[0071] If two images input to the feature extraction network originate from the same ground control point, their feature representations should be similar, resulting in a high similarity value; conversely, they should be separated, resulting in a low similarity value. Based on this principle, automatic identification of corresponding points in images of different resolutions can be achieved by constructing a database of corresponding point samples and training and optimizing the parameters of the feature extraction network.

[0072] In this step, when training the model using the pre-constructed database of corresponding points, conventional deep learning training methods can be employed to train the feature extraction network. The core of the training process lies in optimizing the network parameters to ensure that the feature vectors output by the feature extraction network satisfy the following discriminative conditions: for corresponding point matching samples (positive sample pairs), the similarity between their feature vectors is as high as possible; for non-corresponding point matching samples (negative sample pairs), the similarity between their feature vectors is as low as possible. Through this process, the feature extraction network learns to extract highly discriminative feature representations from the image, thus laying the foundation for subsequent automatic corresponding point recognition.

[0073] Step 3: Preprocessing of the image to be corrected.

[0074] After establishing a sample library of corresponding points and training the feature extraction network model, the image to be corrected is acquired and preprocessed to lay the foundation for accurate matching in the future. Optionally, the image to be corrected can be uploaded to the server backend via a browser frontend.

[0075] The preprocessing of the image to be corrected mainly includes resolution identification, candidate GCP selection, and defining the candidate GCP search window.

[0076] Specifically, the resolution of the acquired image to be corrected is first identified. This invention is limited to supporting three spatial resolution specifications: 0.5 m / pixel, 0.75 m / pixel, and 1.0 m / pixel. By reading the image metadata (such as Ground Sample Distance) of the image to be corrected or by using an image resolution estimation method, the spatial resolution of the image to be corrected is automatically determined, and a GCP template image with the same spatial resolution is selected accordingly.

[0077] Secondly, based on the geographical extent of the image to be corrected, spatial filtering is performed using its vector boundary information and the coordinates of ground control points stored in the GCP database. Only ground control points within the image coverage area are retained, thus selecting candidate GCPs. This step effectively reduces the number of GCPs and decreases the amount of subsequent computation.

[0078] Finally, among the selected candidate GCPs, based on the RPC model or spatial reference system relationship, resolution information, and geographic coordinates of the candidate GCPs in the image to be calibrated, the approximate location range of the candidate GCPs in the image to be calibrated is calculated, and a GCP candidate search window of a regular shape (e.g., rectangle, square) is constructed near this location. This window limits the search space for subsequent fine matching, and finally, a GCP candidate search window is defined for each candidate GCP in the image to be calibrated. Figure 2 Taking the selection of four candidate GCPs as an example, the GCP candidate search window defined for each candidate GCP in the image to be corrected is shown. For the portion of the GCP candidate search window that extends beyond the image edge, black or white can be used to fill in the gaps. This operation establishes an initial correspondence between the candidate GCPs and the image to be corrected, i.e., a pre-matching result. This pre-matching result does not pursue sub-pixel accuracy but rather provides candidate windows and reference ranges for subsequent precise matching.

[0079] Through the above preprocessing, automatic resolution identification of the image to be corrected, rapid screening of candidate GCPs, and reasonable constraints on the search area, i.e., coarse localization, can be achieved, thereby significantly improving the efficiency and stability of subsequent accurate matching of corresponding points.

[0080] Step 4: Automatic matching and visualization of points with the same name.

[0081] After preprocessing the image to be corrected, a trained feature extraction network is used to precisely match candidate GCPs with the defined GCP candidate search windows in the image to be corrected. The main process includes: using the trained feature extraction network to extract the feature vectors of the GCP template image and each local search sub-image within the GCP candidate search window for each candidate GCP. The size of the local search sub-image is the same as that of the GCP template image, 512×512 pixels. For each candidate GCP, the similarity between the feature vectors of its corresponding GCP template image and each local search sub-image is calculated. After the local search sub-image traverses the GCP candidate search window, the local search sub-image corresponding to the maximum similarity is taken as the optimal local search sub-image. The pixel coordinates of the corresponding point of the candidate GCP are calculated based on the position of the optimal local search sub-image. Simultaneously, the calculated pixel coordinates of the corresponding point, along with the corresponding GCP candidate search window and GCP template image, are sent to the browser front-end for visualization and manual review.

[0082] Furthermore, step 4 includes the following steps:

[0083] Step 4.1: Within the GCP candidate search window generated in step 3, slide a fixed-size (512×512 pixels) window with a step size of 1 pixel, and capture a local search sub-image with the same size as the GCP template image corresponding to the candidate GCP, so that the local search sub-image traverses all positions within the GCP candidate search window.

[0084] Step 4.2: For each location, calculate the similarity value between the feature vectors of the GCP template image and the local search sub-image at that location; optionally, use the cosine similarity function to calculate the similarity between the feature vectors. The formula for the cosine similarity function is:

[0085] ;

[0086] in, This represents the input GCP template image, which has the same resolution as the image to be corrected and a size of 512×512 pixels; This indicates that within the GCP candidate search window corresponding to the input GCP template image, the search is conducted using pixel coordinates. The top left corner (i.e., from) The local search sub-image (size 512×512 pixels) is captured at the beginning. The trained feature extraction network is used to encode the image into a feature vector; Similarity is used to measure the degree of similarity between the GCP template image and the local search sub-image.

[0087] Step 4.3: Search within the entire GCP candidate search window to calculate the similarity value. The local search sub-image that reaches the maximum value is taken as the optimal local search sub-image, and its pixel coordinates within the GCP candidate search window are: The optimal local search sub-image is considered to be the best matching point of the candidate GCP corresponding to the GCP template image in the image to be corrected.

[0088] Step 4.4: Combine the pixel coordinates in the image to be corrected using the GCP candidate search window. The pixel coordinates of the optimal local search sub-image within the GCP candidate search window. and the pixel coordinates of the candidate GCP in the GCP template image The pixel coordinates of the corresponding point of the candidate GCP in the image to be corrected are calculated. for:

[0089] ;

[0090] ;

[0091] in:

[0092] The pixel coordinates of the top left corner of the GCP candidate search window in the image to be corrected;

[0093] The pixel coordinates of the optimal local search sub-image obtained within the GCP candidate search window (i.e., the offset with the upper left corner of the GCP candidate search window as the origin).

[0094] The relative pixel coordinates of the candidate GCP within the GCP template image are typically close to (256, 256).

[0095] The pixel coordinates of the corresponding point of the candidate GCP in the image to be corrected are given.

[0096] Finally, the pixel coordinates of the corresponding points of the candidate GCPs are calculated. The precise WGS84 coordinates of the candidate GCP recorded in the GCP database. Establish a one-to-one correspondence, such as Figure 3 As shown:

[0097] ;

[0098] in, The WGS84 coordinates (i.e., the actual geographic coordinates) of the candidate GCP. Longitude Latitude Elevation.

[0099] The final output consists of standardized corresponding point matching pairs, which contain both pixel spatial coordinates and corresponding geographic coordinates, and can be directly used as ground control point inputs for geometric correction. Simultaneously, after completing the automatic and accurate matching of corresponding points, the matching results are displayed visually in the browser front-end (HTML page) for quick manual review and fine-tuning.

[0100] Step 5: Human-computer interaction and sample library closure of matching results.

[0101] In this step, the server backend receives manual review instructions from the browser frontend and updates the corresponding point sample library accordingly. This step not only utilizes a Python-based server backend for feature calculation, similarity evaluation, and data writing, and an HTML or JavaScript-based browser frontend for image rendering and overlay annotation, enabling user interaction and result visualization in a common browser environment, but also eliminates incorrect matches and confirms correct matches. Furthermore, the manual review results are immediately fed back into the sample library, continuously accumulating positive and negative samples as well as difficult positive samples, thus supporting subsequent model retraining and performance iteration.

[0102] At the interface level, the browser front-end displays information for each candidate GCP, including: a partial window of the target large image, a comparison view of the GCP template image, and automatically matched landing points (pixel coordinates). ) and their corresponding WGS84 coordinates Similarity scores and quality indicators (such as correlation peaks and inlier rate) are displayed. Users can perform the following operations on the landing points: First, if a mismatch is determined, it is "removed" with one click, and the pair of samples is written to the negative sample library, recording the image slice, GCP number, resolution level, and contextual features that led to the misjudgment. At this time, the manual review operation instruction received by the server backend is of the type of mismatch removal. The server backend marks the corresponding sample as a negative sample according to the manual review operation instruction, writes it into and updates the corresponding point sample library; Second, if a positioning deviation is determined, sub-pixel fine-tuning is performed in the local magnified view (drag or keyed nudge), and the data is updated accordingly. The sample is then marked as a difficult positive sample and written into the sample library for subsequent targeted retraining in precision-sensitive scenarios. At this point, the type of manual review operation instruction received by the server backend is "matching fine-tuning". The server backend updates the pixel coordinates of the corresponding points of the sample according to the manual review operation instruction, marks the sample as a difficult positive sample, and writes and updates the corresponding point sample library. Third, for confirmed matches, the process can directly enter the geometric fine correction and product generation process. At this point, the type of manual review operation instruction received by the server backend is "corresponding points confirmed". The server backend solidifies the pixel coordinates of the corresponding points confirmed by the manual review operation instruction into standard corresponding point pairs and updates the corresponding point sample library.

[0103] To support the aforementioned human-machine closed loop, the server backend generates auditable records for each operation, including pixel coordinates before and after matching, correction amount (sub-pixel offset), similarity change, operator, timestamp, and resolution label (0.5 / 0.75 / 1 m), and distinguishes sample types (positive samples / negative samples / difficult positive samples). These records are uniformly included in the sample library index for periodic batch retraining. Through direct viewing and review via HTML, this invention achieves a "zero-client installation, cross-platform, low-bandwidth, block-based visualization" annotation experience, significantly reducing the threshold and cost of manual review, while completing the closed loop from automatic matching to sample accumulation with the shortest path. This "Python computing backend + HTML interactive frontend" architecture constitutes a significant advantage of this invention in terms of engineering usability and scalable iteration capabilities.

[0104] Step 6: Geometric fine correction of the image to be corrected.

[0105] Based on the pixel coordinates of the corresponding points confirmed by the manual review operation instructions, the image to be corrected is subjected to geometric fine correction and resampling, and the correction results are returned to the browser front end for visualization.

[0106] After completing the precise matching and manual verification of the corresponding points, this step uses the correspondence between the pixel coordinates of the corresponding points and the geographic coordinates confirmed by the manual verification operation instructions to perform geometric fine correction on the image to be corrected. The basic process includes: establishing the transformation relationship between the image pixel space and the geographic space through GCP and the pixel coordinates of the corresponding points, and reprojecting and resampling the entire image under the constraint of this relationship, so as to obtain an orthophoto consistent with the real geographic coordinate system.

[0107] In practical implementation, firstly, a geometric transformation model is constructed based on the confirmed ground control point matching pairs, i.e., standard corresponding point pairs. For images containing RPC models, bias parameters are estimated and corrected based on the original RPC; for images without RPC, polynomial transformation or thin plate spline (TPS) is used to fit the image-ground relationship. Subsequently, geometric correction and resampling are performed on the entire image. During resampling, methods such as bilinear or cubic convolution interpolation are used to ensure that the image has both high geometric accuracy and good visual quality. The final output is a standardized orthorectified GeoTIFF file of the image to be corrected and a corresponding quality report, which can be directly used in remote sensing mapping, GIS overlay, and geographic analysis applications. Simultaneously, the correction results are returned to the browser frontend for visualization.

[0108] Traditional geometric correction typically relies on large commercial remote sensing software such as PCI and ENVI, which have complex operating environments, high deployment costs, and high hardware requirements. This invention uses only Python code as the computational backend and an HTML page as the front-end interface for user interaction and batch triggering. The Python backend handles ground control point fitting, model calculation, and image reprojection, while the HTML frontend provides a "one-click batch correction" button and progress display. Users can complete the entire geometric correction process using only a general-purpose computer and browser, without needing to install additional large software or specialized equipment.

[0109] Through the above steps, this invention achieves a complete process from ground control point matching to geometric precision correction. Its main advantages include: low deployment and maintenance costs, lightweight and flexible system; high processing efficiency, supporting rapid correction of batch images; and convenient user operation, requiring only one click in a browser to complete complex geometric correction tasks. The lightweight architecture of this invention not only significantly lowers the technical threshold but also enhances its scalability and application value in large-scale remote sensing image processing scenarios. Simultaneously, this invention achieves significant comprehensive benefits: in terms of efficiency, automated matching reduces the time for human-machine collaborative ground control point selection by approximately 80%; in terms of economy, it completely replaces large-scale commercial remote sensing software, effectively reducing usage costs; in terms of applicability, its lightweight architecture allows for smooth operation on ordinary desktop computers, reducing dependence on dedicated hardware; and it supports manual fine-tuning, ensuring correction accuracy, with a root mean square error (RMSE) within 0.2 pixels, achieving both efficient processing and high-precision output.

[0110] This invention also proposes a lightweight remote sensing image geometric fine correction system based on a B / S architecture for implementing the methods described in the foregoing embodiments, such as... Figure 4 As shown, the system adopts a B / S architecture consisting of a browser front-end and a server back-end, specifically including:

[0111] Deployed on the server backend:

[0112] The GCP management module is used to build the GCP database. The GCP database stores the geographic coordinates of several GCPs, and the GCP template map and its spatial resolution under three different spatial resolution specifications of 0.5 meters / pixel, 0.75 meters / pixel and 1.0 meters / pixel for each GCP. The size of each GCP template map is 512×512 pixels.

[0113] The sample library construction and training module is used to construct a sample library of corresponding points based on the GCP database, and to train and optimize the parameters of the feature extraction network using the sample library of corresponding points.

[0114] The preprocessing module is used to acquire the image to be corrected and to preprocess the image to be corrected, including identifying the resolution, screening candidate GCPs, and defining a GCP candidate search window in the image to be corrected for each candidate GCP.

[0115] The corresponding point automatic matching module is used to extract the feature vectors of the GCP template image and each local search sub-image with a size of 512×512 pixels in the GCP candidate search window for each candidate GCP using the trained feature extraction network. It also calculates the similarity between the feature vectors of the GCP template image and each local search sub-image. After the local search sub-image traverses the GCP candidate search window, the pixel coordinates of the corresponding point of the candidate GCP are calculated based on the position of the optimal local search sub-image corresponding to the maximum similarity.

[0116] The sample library update module is used to receive manual review operation instructions and update the sample library of corresponding points according to the manual review operation instructions.

[0117] The geometric correction module is used to perform geometric fine correction and resampling on the image to be corrected based on the pixel coordinates of the corresponding points confirmed by the manual review operation instructions, and return the correction results to the browser front end for visualization.

[0118] And, running in the browser frontend:

[0119] The human-computer interaction and visualization module is used to visualize the pixel coordinates of the corresponding points output by the automatic matching module and their corresponding GCP candidate search window and GCP template image through a visualization interface. It also receives interactive operations triggered by the user based on the visualization interface (such as elimination trigger, sub-pixel fine-tuning trigger, confirmation trigger), and generates and outputs the corresponding manual review operation instructions to the sample library update module.

[0120] In this embodiment, the modules are configured to work collaboratively to implement the methods described in the foregoing embodiments. The specific working process of each module has been described in detail in the foregoing method embodiments, and will not be repeated here.

[0121] This invention provides a lightweight remote sensing image geometric correction visualization system based on a B / S architecture. The system uses Python as the computational backend and HTML as the front-end interactive interface. Its overall architecture is lightweight and efficient, capable of running on general-purpose computers and browsers. Unlike traditional methods that rely on large commercial software, this system can automatically identify corresponding points in remote sensing images, perform manual review, geometric correction, and output results without the need for complex platforms such as PCI or ArcGIS. The system is characterized by easy deployment, low maintenance costs, and flexible upgrades, significantly lowering the barrier to entry and supporting closed-loop updates of the sample database to continuously improve recognition accuracy.

[0122] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0123] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A lightweight remote sensing image geometric correction method based on B / S architecture, characterized in that, The method is executed in a B / S architecture consisting of a browser front-end and a server back-end, wherein the server back-end performs the following steps: Step 1: Construct a GCP database, which stores the geographic coordinates of several GCPs, and the corresponding 0.5 meters / pixel for each GCP. GCP template images and their spatial resolutions under three different spatial resolution specifications: 0.75 m / pixel and 1.0 m / pixel, and the size of the GCP template image is 512×512 pixels; Step 2: Construct a corresponding point sample library based on the GCP database, and use the corresponding point sample library to train and optimize the parameters of the feature extraction network; Step 3: Acquire the image to be corrected and preprocess the image to be corrected, including identifying the resolution, filtering candidate GCPs, and defining a GCP candidate search window for each candidate GCP in the image to be corrected; Step 4: Using the trained feature extraction network, extract the feature vectors of the GCP template image and each local search sub-image with a size of 512×512 pixels in the GCP candidate search window for each candidate GCP. Calculate the similarity between the feature vectors of the GCP template image and each local search sub-image. After the local search sub-image traverses the GCP candidate search window, calculate the pixel coordinates of the corresponding point of the candidate GCP based on the position of the optimal local search sub-image corresponding to the maximum similarity. Send the pixel coordinates of the corresponding point and the corresponding GCP candidate search window and GCP template image to the browser front-end for visualization. Step 5: Receive manual review instructions from the browser front-end, and update the same-name point sample library according to the manual review instructions; Step 6: Based on the pixel coordinates of the corresponding points confirmed by the manual review operation command, perform geometric fine correction and resampling on the image to be corrected, and return the correction results to the browser front end for visualization display.

2. The lightweight remote sensing image geometric correction method based on B / S architecture according to claim 1, characterized in that, In step 1, building the GCP database includes the following steps: Acquire high-resolution imagery and the geographic coordinates of each GCP on it; The high-resolution image was downsampled to a spatial resolution of 0.1 m / pixel, and a GCP original reference sub-image with a size of 512×512 pixels was cropped with GCP as the center. Spatial resolution transformation is performed on each of the original GCP reference sub-images to generate 0.5 m / pixel. GCP template images with three different spatial resolutions: 0.75 meters / pixel and 1.0 meters / pixel; Each GCP template diagram and its related information are stored in a unified database to construct a GCP database.

3. The lightweight remote sensing image geometric correction method based on B / S architecture according to claim 1, characterized in that, In step 2, constructing the sample library of corresponding points includes the following steps: Select remote sensing images containing at least one GCP and with spatial resolutions of 0.5 m / pixel, 0.75 m / pixel, and 1.0 m / pixel, respectively, as the target images; On the target images at each spatial resolution, local area images of size 512×512 pixels are cropped out with the geographic coordinates of the GCPs included therein as the center. For each spatial resolution specification, an image pair consisting of the GCP template image of each GCP and its corresponding local area image is used as a corresponding point matching sample. All the corresponding point matching samples are stored in the database to construct a corresponding point sample library.

4. The lightweight remote sensing image geometric correction method based on B / S architecture according to claim 3, characterized in that, Each sample in the same point sample library contains the following information: (1) GCP template diagram and its spatial resolution; (2) Local area image corresponding to the GCP template image; (3) The relationship between the corresponding points established at the pixel level between the GCP template image and the corresponding local area image; (4) Metadata information.

5. The lightweight remote sensing image geometric correction method based on B / S architecture according to claim 3 or 4, characterized in that, In step 2, training and optimizing the parameters of the feature extraction network using the same point sample library includes the following steps: The corresponding point matching sample is input into the feature extraction network, and the feature extraction network extracts feature vectors from the GCP template map and the local area image in the corresponding point matching sample, respectively. The degree of similarity between the two feature vectors is calculated using a similarity metric function; The algorithm iterates to optimize the parameters of the feature extraction network, maximizing the similarity between images that are matched from the same point and minimizing the similarity between images that are not matched from the same point.

6. The lightweight remote sensing image geometric correction method based on B / S architecture according to any one of claims 1 to 4, characterized in that, Step 4 includes the following steps: Step 4.1: Within the GCP candidate search window, slide the window pixel by pixel to capture the local search sub-image, so that the local search sub-image traverses all positions within the GCP candidate search window; Step 4.2: For each location, calculate the similarity value between the feature vector of the GCP template image and the feature vector of the local search sub-image at that location; Step 4.3: Determine the optimal local search sub-image that maximizes the similarity value, whose pixel coordinates within the GCP candidate search window are: ; Step 4.4: Combine the pixel coordinates in the image to be corrected using the GCP candidate search window. Pixel coordinates and the pixel coordinates of the candidate GCP in the GCP template image The pixel coordinates of the corresponding point of the candidate GCP in the image to be corrected are calculated. ,in , .

7. The lightweight remote sensing image geometric correction method based on B / S architecture according to claim 6, characterized in that, The cosine similarity function is used to calculate the similarity between feature vectors. The formula is as follows: ； in, The input GCP template diagram; To search for GCP candidates within the GCP candidate search window corresponding to the input GCP template image, using pixel coordinates. This is a local search sub-image captured starting from the top left corner; For a well-trained feature extraction network; This is the L2 norm of the feature vectors extracted by the feature extraction network.

8. The lightweight remote sensing image geometric correction method based on B / S architecture according to any one of claims 1 to 4, characterized in that, In step 5, the following operations are performed according to the type of the manual review operation instruction: If the type of the manual review operation instruction is mismatch elimination, then the corresponding sample is marked as a negative sample according to the manual review operation instruction, and written into and updated the same point sample library; If the type of the manual review operation instruction is matching fine-tuning, then update the pixel coordinates of the corresponding sample's corresponding point according to the manual review operation instruction, and mark the sample as a difficult positive sample, write it into and update the corresponding point sample library; If the type of the manual verification operation instruction is "Correct confirmation of corresponding points", then the pixel coordinates of the corresponding points confirmed by the manual verification operation instruction are fixed as standard corresponding point pairs, and the corresponding point sample library is updated.

9. The lightweight remote sensing image geometric correction method based on B / S architecture according to any one of claims 1 to 4, characterized in that, The browser front-end uses HTML or JavaScript to enable user interaction and result visualization in a general browser environment; the server back-end uses Python to implement computing tasks.

10. A lightweight remote sensing image geometric correction system based on a B / S architecture for implementing the method as described in any one of claims 1 to 9, characterized in that, The system adopts a B / S architecture consisting of a browser front-end and a server back-end, including: Deployed on the backend of the server: The GCP management module is used to build a GCP database, which stores the geographic coordinates of several GCPs and the corresponding 0.5 meters / pixel for each GCP. GCP template images and their spatial resolutions under three different spatial resolution specifications: 0.75 m / pixel and 1.0 m / pixel, and the size of the GCP template image is 512×512 pixels; The sample library construction and training module is used to construct a sample library of corresponding points based on the GCP database, and to train and optimize the parameters of the feature extraction network using the sample library of corresponding points. The preprocessing module is used to acquire the image to be corrected and to preprocess the image to be corrected, including identifying the resolution, filtering candidate GCPs, and defining a GCP candidate search window for each candidate GCP in the image to be corrected. The corresponding point automatic matching module is used to extract the feature vectors of each candidate GCP template image and each local search sub-image with a size of 512×512 pixels in the GCP candidate search window using a trained feature extraction network, and calculate the similarity between the feature vectors of the GCP template image and each local search sub-image. After the local search sub-image traverses the GCP candidate search window, the pixel coordinates of the corresponding point of the candidate GCP are calculated based on the position of the optimal local search sub-image corresponding to the maximum similarity. The sample library update module is used to receive manual review operation instructions and update the sample library of the same name points according to the manual review operation instructions. The geometric correction module is used to perform geometric fine correction and resampling on the image to be corrected based on the pixel coordinates of the corresponding points confirmed by the manual review operation instructions, and return the correction results to the browser front end for visualization display. And, running in the browser front end: The human-computer interaction and visualization module is used to visualize the pixel coordinates of the corresponding points output by the automatic matching module and their corresponding GCP candidate search windows and GCP template diagrams through a visualization interface, and to receive interactive operations triggered by the user based on the visualization interface, and generate and output the corresponding manual review operation instructions to the sample library update module.

Images