Vector geographic target overall extraction method and device for large-breadth remote sensing image

Abstract

The invention provides a vector geographic target overall extraction method and device for a large-format remote sensing image. The method comprises the following steps: a context sensing segmentation stage: processing the whole remote sensing image by adopting a context attention network, extracting context information through a local-to-global attention mechanism, and generating a segmentation mask of a target; in the mask edge reconstruction stage, a target contour is extracted from the segmented mask, and a regular polygon sequence with uniformly distributed vertexes is generated through polygon simplification and fixed distance interpolation reconstruction; in the polygon sequence tracking stage, the regularized polygon sequence is input into a polygon sequence tracker, and a complete vectorized geographic target is output through position offset correction and vertex classification; wherein in the three stages, end-to-end whole-scene processing is carried out on a large-format remote sensing image, and blocking operation is not needed. The method aims at overcoming the limitation of a traditional method and providing a complete and coherent vector result for high-precision geographic information extraction.

Description

Technical Field

[0001] The present invention belongs to the field of remote sensing image processing and computer vision, and in particular relates to a technical solution for automatically and holistically extracting vectorized geographic targets (such as buildings, water bodies, roads, etc.) from large-scale remote sensing images. Background Art

[0002] Vector maps accurately represent the Earth's surface and are essential foundational data for numerous applications, including geographic information systems (GIS), autonomous driving navigation, land and resources monitoring, and urban planning and management. Automatically extracting high-precision vector maps using remote sensing imagery (RSI) has become a key technology for acquiring and updating geographic information due to its wide coverage, short update cycles, and high cost-effectiveness, and has achieved significant technological progress in recent years.

[0003] With the rapid development of aerospace remote sensing technology, the spatial resolution of remote sensing images acquired by Earth observation systems has continued to increase, and the geographic coverage of single images has also expanded, resulting in a growing number of large-scale, high-resolution remote sensing image data sources. These large-scale images provide rich information for high-precision vector mapping of geographic targets. However, existing automated or semi-automated vector map extraction methods often struggle with large-scale remote sensing images, such as those exceeding 10,000 × 10,000 pixels.

[0004] Specifically, current mainstream vectorization algorithms, whether based on semantic segmentation post-processing or direct geometric feature prediction, are often limited by computing resources such as graphics processing unit (GPU) memory capacity. Consequently, their model design and training are often based on smaller image blocks (e.g., 512×512). When processing large-scale remote sensing imagery, these methods generally employ a simple "divide and conquer" strategy: first, the large image is cropped into several small, non-overlapping or slightly overlapping blocks. Vectorization is then performed independently on each block. Finally, the processed results from each block are combined to form a vector map covering the entire area.

[0005] However, this patch-based processing strategy has inherent limitations and poses severe technical challenges. First, contextual information is lost. Independently processing image patches severs the intrinsic spatial connection between geographic targets and their surroundings, causing the model to lose key contextual information necessary to understand the target's attributes. For example, in the absence of sufficient surrounding environmental information, the model may find it difficult to accurately distinguish between objects that appear similar but have different semantics, such as misjudging a building roof as a parking lot, or incorrectly interrupting a road at an image patch boundary. Second, the results are fragmented. Because each image patch is processed independently, the vector representations of geographic targets located at the image patch boundary (such as roads, rivers, or large buildings that cross the boundary) often break, misalign, or have discontinuous geometric shapes at the splicing points, seriously damaging the topological consistency and geometric integrity of the vector results. Complex post-processing steps are required to repair them, and the results are difficult to guarantee.

[0006] Furthermore, large-scale remote sensing images contain a wide variety of geographic objects, with diverse shapes and scales. For example, long, winding rivers and road networks differ greatly from scattered small buildings in terms of spatial scale and geometric complexity. This inherent scale diversity not only exacerbates the fragmentation problem caused by image block processing strategies, but also places higher demands on the implementation of a vectorization model that can uniformly and robustly represent different types of geographic objects. Although existing research has made progress in the vectorization of specific categories of objects (such as building extraction and road extraction), and a small amount of work has attempted to uniformly extract multiple categories of geographic objects, most of them have not yet broken away from the paradigm of image block processing and have failed to fundamentally solve the problems of context loss and result fragmentation faced when processing large-scale remote sensing images.

[0007] In summary, existing technologies for extracting vectorized geographic objects from large, high-resolution remote sensing imagery face bottlenecks such as insufficient use of contextual information, fragmented output, and difficulty adapting to diverse object scales. Therefore, there is an urgent need to research and develop innovative methods that can directly and holistically process large remote sensing imagery, effectively utilize global contextual information, and generate geometrically complete and topologically correct vectorized geographic objects. Summary of the Invention

[0008] This invention aims to address several technical bottlenecks in existing remote sensing image vectorization methods, as mentioned in the background, caused by the widespread use of block-based processing strategies when processing large-scale, high-resolution imagery. These bottlenecks primarily include: loss of critical contextual information, resulting in reduced object recognition accuracy; fragmentation and geometric discontinuities in vectorized results at block boundaries, which undermines the integrity of geographic objects; and difficulty in effectively adapting to and uniformly representing the significant scale variations of different geographic object categories in large-scale imagery.

[0009] In view of these shortcomings, the present invention provides a method for directly and holistically extracting vectorized geographic targets from large-scale remote sensing images.

[0010] To achieve the above-mentioned purpose, the technical solution of the present invention is a method for extracting vector geographic targets from large-scale remote sensing images, comprising: In the context-aware segmentation stage, a contextual attention network is used to process the entire remote sensing image, extracting contextual information through a local-to-global attention mechanism to generate a segmentation mask for the target. In the mask edge reconstruction stage, the target outline is extracted from the segmentation mask and reconstructed through polygon simplification and fixed distance interpolation to generate a regularized polygon sequence with uniform vertex distribution; In the polygon sequence tracking stage, the regularized polygon sequence is input into the polygon sequence tracker, and the complete vectorized geographic target is output through position offset correction and vertex classification; Three of the stages perform end-to-end scene processing on large-scale remote sensing images without the need for segmentation.

[0011] Moreover, the working process of the context-aware segmentation stage includes the following processing, Construct a multi-scale pyramid of input images; Use sliding windows to synchronously extract image blocks of various scales; Extract features through an encoder with shared weights; The underlying high-resolution features are used as queries, multi-scale features are used as keys, and the attention mechanism is used to fuse contextual information. After concatenating the fusion features, the decoder outputs the segmentation mask.

[0012] Furthermore, the working process of the mask edge reconstruction stage includes extracting an initial contour point set from the segmentation mask, simplifying the contour point set using a tolerance parameter, and performing linear interpolation at a fixed distance along each edge of the simplified polygon to generate a regularized polygon sequence.

[0013] Moreover, in the training of mask edge reconstruction, the reconstructed polygon sequence is matched with the true polygon, and vertex or non-vertex labels are assigned to the points in the sequence; and the true position data corresponding to the reconstructed sequence points are generated.

[0014] Furthermore, when position offset correction is performed in the polygon sequence tracking stage, image features of each point in the regularized polygon sequence are extracted, and based on the image features and point coordinates, position offsets are predicted and coordinates are corrected using a sequence model.

[0015] Furthermore, when performing vertex classification in the polygon sequence tracking stage, the local geometric angle features of each point in the corrected polygon sequence are calculated, the coordinates, image features and angle features are fused, the vertex probability is predicted, and the final vertex set is determined.

[0016] Moreover, a combined loss function is used when training the polygon sequence tracker, including the regression loss of position offset, the cross entropy loss of vertex classification, and the penalty loss based on geometric angles. The three types of losses are weighted and summed by preset weight coefficients.

[0017] On the other hand, the present invention provides an electronic device comprising a memory, a processor and a computer program stored in the memory, wherein when the processor executes the program, the method for overall extraction of vector geographic targets for large-scale remote sensing images as described above is implemented.

[0018] On the other hand, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned method for overall extraction of vector geographic targets for large-scale remote sensing images.

[0019] On the other hand, the present invention provides a computer program product, including a computer program, which, when executed by a processor, implements the above-mentioned method for overall extraction of vector geographic targets for large-scale remote sensing images.

[0020] In conclusion, in view of the problem that existing remote sensing image vectorization methods are usually limited to processing small image blocks when processing high-resolution and large-scale images, resulting in loss of contextual information and fragmented vectorization results, the present invention discloses a whole-scene learning technology for unified vector generation of large-scale remote sensing image objects. Different from the traditional block processing strategy, the present invention directly processes the whole-scene image to eliminate stitching errors. The present invention first adopts a contextual attention network, and uses its local to global attention mechanism to effectively capture large-scale contextual dependencies, significantly enhancing the segmentation effect of large-scale remote sensing images; then, through a robust vectorization process, a mask edge reconstructor is used to perform polygon reconstruction, and combined with a polygon sequence tracker for vertex tracking, the overall and continuous vectorization of geographic targets is achieved. The present invention aims to overcome the limitations of traditional methods and provide complete and coherent vector results for high-precision geographic information extraction.

[0021] The present invention provides a novel, holistic vectorization framework suitable for large-scale remote sensing images, which effectively solves the key challenges faced by existing technologies and significantly improves the integrity, accuracy and automation level of vectorized geographic target extraction. It is particularly suitable for fields that require high-precision vector data, such as land surveys and smart city three-dimensional modeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a flow chart of a method for extracting vector geographic targets from large-scale remote sensing images according to an embodiment of the present invention.

[0023] Figure 2 2 is a schematic diagram of the principle of a mask edge reconstructor according to an embodiment of the present invention.

[0024] Figure 3 yes Figure 1 A local zoomed-in view of the context-aware network in

[15] .

[0025] Figure 4 yes Figure 1 A zoomed-in view of the polygon sequence tracker in [1]. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0027] This paper proposes a whole-scene learning technology for unified vector generation of objects in large-scale remote sensing imagery, supporting unified tracking of large-scale remote sensing image vectorization. The core concept of this approach is to overcome the limitations of existing block-based processing strategies. By combining context-aware segmentation, regularized reconstruction of polygon outlines, and refined tracking and vertex recognition based on sequence processing, it achieves end-to-end, high-precision vectorization extraction of geographic objects in large-scale imagery. This method can be implemented through computer software programming and run on hardware platforms equipped with computing units such as graphics processors.

[0028] Figure 1 This article demonstrates the overall processing flow of one embodiment of the present invention. This embodiment provides a method for extracting vector geographic objects from large-scale remote sensing imagery. The method comprises three core modules / stages: a Context Attention Network (CAN) module, which performs context-aware image segmentation; a Mask Edge Reconstructor (MCR) module, which converts the segmentation results into regularized polygon representations; and a Polygon Sequence Tracker (PST) module, which refines polygons and identifies precise vertices, ultimately outputting vectorized geographic objects. The specific implementation of each module / stage is detailed below.

[0029] Step 1: Processing the large-scale remote sensing image to generate a context-aware segmentation result, wherein the processing utilizes a context attention network that extracts context information from the large-scale remote sensing image using a local-to-global attention mechanism; The Contextual Attention Network (CAN) module is designed to perform pixel-level semantic segmentation on large-scale input remote sensing images while effectively capturing and leveraging the image's global context to generate more complete and accurate segmentation masks than traditional block-based processing methods. First, a multi-scale image pyramid is constructed for the input large-scale remote sensing image by applying different downsampling rates. A unified image encoder (which may employ a weight-sharing strategy) is used to extract feature maps for each scale in the pyramid. A contextual attention mechanism is employed to achieve local-to-global feature fusion. Specifically, the feature map at the bottom (highest resolution) of the pyramid serves as the query, and feature maps at each scale (including its own and other scales) serve as the key and value. Enhanced features at each scale are generated through attention calculations that incorporate contextual information. Finally, the bottom feature map and all fused enhanced feature maps are concatenated and fed into a semantic segmentation decoder to generate a relatively complete pixel-level object segmentation mask that reflects global contextual information.

[0030] See also Figure 3 , the specific implementation of the embodiment includes the following processes: 1.1 Multi-scale image pyramid construction: A multi-scale image pyramid is constructed by applying different downsampling rates to the large-scale remote sensing image; Given a large-scale remote sensing image , first applying a set of different downsampling rates (generally and ) to construct a multi-scale image pyramid. The bottom layer of the pyramid is the original resolution image (or the downsampling rate is 1), and the top layer is the lowest resolution image.

[0031] 1.2 Pyramid Slicing and Feature Extraction: The multi-scale image pyramid is sliced ​​using a sliding window of uniform size to obtain multiple groups of image blocks of the same size but different scales. The image blocks of each scale are processed using a shared weight image encoder to extract feature maps. like Figure 1 As shown, a fixed size (e.g. For the same spatial position on the pyramid, a set of image blocks with the same size but different resolutions can be obtained. ,in The corresponding downsampling rate is The layer, The image blocks are selected in a way that ensures that the low-resolution blocks The field of view includes high-resolution blocks Then, a shared weighted frozen image encoder (weights do not participate in back propagation) is used to process these image blocks separately and extract the corresponding feature maps. Only the bottom layer (n=1) encoder is trained, and other layers share their weights for greater efficiency.

[0032] 1.3 Contextual Attention Fusion: Using the contextual attention mechanism to fuse the underlying feature map with other scale feature maps In order to fuse multi-scale information and capture long-distance dependencies, a context attention mechanism is adopted. As a query ), and each scale ( , including itself) feature map Also serves as a key (Key, ) and Value ). Through the standard scaled dot product attention calculation, the enhanced features of each scale that incorporate contextual information are obtained :

[0033] in is the dimension of the key vector. This approach allows features at each scale to focus on local details (from ) and global structures (from and other scales).

[0034] 1.4 Feature concatenation and decoding: The underlying feature map and all fused feature maps are concatenated and input into the segmentation decoder to generate the context-aware segmentation result.

[0035] The underlying original feature map Combined with all enhanced feature maps obtained through attention fusion Concatenate along the channel dimension:

[0036] The obtained fusion features This contains rich multi-scale contextual information. Finally, the fused features are input into a semantic segmentation decoder (for example, using an UperNet structure with powerful feature fusion capabilities). Through upsampling and convolution operations, a pixel-level semantic segmentation probability map is finally generated, and then the target segmentation mask is obtained.

[0037] When training the above CAN module, it is recommended to use the standard cross-entropy loss function (Cross-EntropyLoss) as the supervision signal:

[0038] in, is the semantic segmentation loss, is the total number of pixels, is the number of categories, It's a pixel Belong to category The true value label (0 or 1), is the model prediction pixel Belong to category probability.

[0039] Step 2: Processing the segmentation results to generate a regularized polygonal representation, the processing is achieved by a mask edge reconstructor, which extracts the polygon outline, simplifies it, and reconstructs the polygon by interpolating along the simplified edges at a fixed distance; The Mask Edge Reconstructor (MCR) module receives the segmentation mask from the CAN module and aims to extract the contour of the target and convert it into a polygon sequence with more uniform vertex distribution and more regular structure, so as to facilitate the processing and training of the subsequent PST module. From the segmentation mask generated in the previous stage, the initial contour of the target is extracted using a contour extraction algorithm (such as TC89-KCOS), resulting in a sequence of dense and possibly irregular points. Apply polygon simplification algorithms (such as Douglas-Peucker algorithm) according to the preset tolerance For the initial contour sequence Simplify to obtain a simplified polygon with a reduced number of vertices To make the polygon representation more regular, simplify the polygon Each edge of Perform linear interpolation to generate a new polygon sequence with relatively uniform vertex distribution This sequence As input to the next stage. To generate training data, the reconstructed polygons The corresponding true polygon Match and confirm Which points are vertices (labels ) which are not (label ), and generate a Reconstructed true polygons with the same number of points and one-to-one correspondence .

[0040] See also Figure 2, the specific implementation of the embodiment includes the following processes: 2.1 Initial contour extraction: Extract the initial contour point set from the segmentation result For each connected component of the target in the segmentation mask, a standard contour finding algorithm (e.g., an implementation of the Teh-Chin algorithm, such as TC89-KCOS) is used to extract its boundary contours, obtaining an initial contour point set consisting of pixel coordinates. ,in , dot mark , is the number of contour points, are the point coordinates. These points are usually very dense and irregularly arranged.

[0041] 2.2 Contour simplification: Apply the polygon simplification algorithm and the preset tolerance to process the initial contour point set to obtain a simplified vertex sequence In order to reduce redundant points and preserve the main shape of the contour, a polygon simplification algorithm (e.g., the classic Douglas-Peucker algorithm) is used to process the initial contour point set. . Set a tolerance parameter The algorithm will remove those points whose distance to the line segment connecting the previous and next key points is less than points, resulting in a simplified polygon with a greatly reduced number of vertices ,in and .

[0042] 2.3 Fixed-distance interpolation reconstruction: along the simplified vertex sequence Each edge of , according to the preset fixed distance Interpolate to generate A regular polygon sequence of points .

[0043] In order to make the polygon's vertex distribution more uniform and facilitate subsequent sequence-based model processing, MCR simplifies the polygon Each edge of (Notice Set a fixed interpolation distance , insert evenly on each edge Points, is the number of insertion points. Combined with all interpolation points in sequence to obtain the final reconstructed polygon sequence , where the total number of points , which includes the vertex This sequence Will be used as input to the PST module.

[0044] The present invention further provides a matching and label generation implementation scheme for the training phase of the PST module: the reconstructed regular polygon sequence The corresponding geographic target ground truth polygon Perform two-way matching for Each point in Assign labels to vertices or non-vertices , and generate Reconstructed ground truth polygons with one-to-one correspondence , that is, providing the true value position data corresponding to the reconstructed sequence points for training in subsequent steps. Include Vertex .

[0045] During the training phase, when training the PST module, it is necessary to reconstruct the polygon sequence To provide supervision information (i.e. which points are real vertices and their precise locations), MCR performs an additional matching step during the training phase: Reconstructing polygons The corresponding true polygon (usually manually annotated or obtained from existing vector data) for matching. A two-way matching mechanism is used: for the true value Each vertex in ,exist Find the point with the closest Euclidean distance ,in represents the two-norm, Represents the reconstructed polygon points.

[0046] Collect all matched Point in Since there may be multiple vertices Match to the same vertex In the case of , take the unique set of these matching points and sort them by Sort the original index in , and get ( ) points, which are initially considered to be The candidate vertices corresponding to the true vertices in .

[0047] for Each point in Assign a binary label :if is the above One of the candidate vertices, then ;otherwise . Forming a tag sequence .

[0048] In order to provide PST with a supervision target for position regression, it is necessary to generate a "Reconstructed true value polygon" with the same number of points and one-to-one correspondence between points The method is: Any two consecutive points marked as vertices in and ( ), there is non-vertices. Find the two vertices in the true value The corresponding original vertex in and , then on the true value side Upper uniform interpolation Points. Combine all true value vertices and all interpolation points in order to get It can be proved that .so, Each point in There is a corresponding truth value position , and a vertex label like Figure 2 The num of the training samples for both the upper and lower images is 4.

[0049] In the prediction phase, when the model is inferring (predicting), the MCR module executes steps 2.1 to 2.3 accordingly, i.e., extracting contours, simplifying, and reconstructing, and directly outputs a regularized polygon sequence. , no matching with the true value is performed.

[0050] Step 3, processing the regularized polygon representation to generate a final vectorized geographic target, wherein the processing adopts a polygon sequence tracker, which treats polygons as sequences, corrects the position offsets of points in the sequence, and performs vertex classification to identify accurate vertices.

[0051] The Polygon Sequence Tracker (PST) module receives the regularized polygon sequence from the MCR , treat it as a sequence of points, perform refined geometric position correction and accurate vertex recognition, and finally output high-quality vectorized geographic targets. Treat it as a sequence of points and process it. , extract the depth image features at the corresponding position of the original image (You can use the encoder trained in the first stage and freeze its weights). and image features Combined, input into an offset regressor based on a sequence processing model (such as Transformer) to predict the position correction offset of each point Apply the offset to the original coordinates to get the corrected point This step can be repeated multiple times to refine the position. Based on the corrected point sequence , calculate each point For example, calculate the local geometric angle characteristics at The distance between them is ( Can take multiple values, such as 1, 2, 3) The angle formed by the neighboring points (For example, the formula ). The corrected coordinates , corresponding image features And the calculated angle characteristics The information is integrated and input into a vertex predictor based on a sequence processing model (such as Transformer) for each point in the sequence. Output a probability of being a real vertex According to the predicted vertex probability (For example, by setting a threshold or directly selecting the point with the highest probability as the vertex), the precise vertex sequence of the geographic target is finally determined, thereby outputting a complete vectorized representation.

[0052] When training the PST model, a combined loss function is used, which may include a regression loss (such as Smooth L1 Loss) for supervising offset prediction, a classification loss (such as binary cross entropy loss) for supervising vertex classification, and an optional angle penalty loss to encourage sharp corners to be identified as vertices.

[0053] See also Figure 4 , the specific implementation of the embodiment includes the following processes: 3.1 Image feature extraction: regularized polygon sequence Each point in , extract the corresponding image features ; For the input sequence Every point in , using the image encoder of the trained CAN module (the weight is usually frozen and not updated at this time), extract the local image features around the corresponding position of the point in the original image and obtain the feature vector .

[0054] 3.2 Position offset correction: Image features With point coordinates Combined, input to the Transformer-based offset regressor to predict the offset correction for each point ; Apply the predicted offset correction to the corresponding point to obtain the corrected point coordinates , forming a modified polygon sequence ; Optionally, repeat this process for multiple rounds of iterative correction.

[0055] Specifically, in order to further correct the possible position deviation caused by MCR reconstruction, PST uses an offset regressor based on a sequence processing model (such as the Transformer encoder-decoder structure). Coordinates With its image features After concatenation or fusion (forming input ) is fed into the regressor to predict a position correction offset Add the predicted offset to the original coordinates to get the corrected coordinates .

[0056] Iterative correction: In order to capture more precise geometric adjustments, we can refer to the ideas of DeepSnake and other works and perform multiple rounds of iterative correction. Sequence after round correction As input, predict Wheel offset , then update the sequence .

[0057] 3.3 Angular feature calculation: based on the corrected polygon sequence , calculate each point The preset distance before and after (For example ) formed by the adjacent points ; Recognizing that vertices usually correspond to sharp angles or corners on polygon outlines, PST explicitly calculates and utilizes angle features to assist vertex recognition. , for each point , calculate the different distances between it and the front and back (For example, )'s neighboring points The calculation formula is as follows:

[0058] in and It is a connection point Before and after The vector of neighbors (be careful to handle sequence boundaries, for example using periodic boundary conditions , mod represents the remainder operation, j satisfies the periodicity condition to adapt to the angle calculation of the first and last points). The angle value obtained exist Within the range, reflecting the point These angles can be further transformed to polar space or used directly as features.

[0059] 3.4 Vertex classification prediction: Corrected point coordinates , corresponding image features And the calculated angle characteristics Combined, input to the Transformer-based vertex predictor, output each point The probability of a vertex ; Specifically, another vertex predictor based on a sequence processing model (such as Transformer) is used. The coordinates and corresponding image features (which may also need to be updated) And the calculated angle characteristics Multimodal information is integrated to form input , and then fed into the vertex predictor. The predictor is for each point in the sequence Output a probability value , indicating the possibility that the point is a real vertex.

[0060] 3.5 Final vector output: according to the vertex probability Determine the final vertex set and combine it with the corrected point coordinates Output the vectorized geographic object.

[0061] According to the predicted vertex probability sequence , different strategies can be used to determine the final vertex set. For example, a probability threshold can be set , all point Alternatively, methods such as non-maximum suppression (NMS) can be applied to select the points with the highest local probability as vertices. The selected vertex sequence constitutes the final output vectorized geographic target.

[0062] In the embodiment, when training the PST module, a composite loss function comprising multiple components is used. To optimize both position correction and vertex classification tasks, for example, including offset regression loss , vertex classification loss and angle penalty loss The combined loss function is: Bias regression loss : The offset of the supervised prediction As close as possible and The actual offset between (or If the supervision target is the corrected position), Smooth L1 loss can be used:

[0063] Vertex classification loss : Vertex probability of supervised prediction Close to the true value label generated by MCR . Binary Cross-Entropy (BCE) loss can be used:

[0064] Angle Penalty Loss : In order to take advantage of the geometric prior (i.e. vertices usually correspond to small angles), an angle penalty term can be introduced to encourage the model to have angles less than a certain threshold The point prediction has a higher vertex probability (if its true value label ), and predict lower vertex probabilities for points with angles greater than the threshold (if their true value labels ).

[0065] The weighted sum of the above losses gives the total loss function:

[0066] The preset weight coefficient is a hyperparameter used to balance the importance of different tasks. The PST model is trained by minimizing this total loss.

[0067] In specific implementation, the method of the present invention can also be implemented as a system. The system may include: Image input and preprocessing unit: responsible for receiving large-scale remote sensing images and performing multi-scale pyramid construction and slicing operations.

[0068] Context-aware segmentation unit (implements CAN functionality): Contains an image encoder with shared weights, a contextual attention module, and a segmentation decoder to generate the initial segmentation mask of the object.

[0069] Contour reconstruction unit (implements MCR functionality): Responsible for extracting and simplifying contours from the segmentation mask, reconstructing them using fixed-distance interpolation, and outputting a regularized polygon sequence. In training mode, it also performs ground truth matching and label generation.

[0070] Sequence tracking and vertex recognition unit (implementing PST function): Contains image feature extractor, offset regressor, angle feature calculator and vertex predictor, responsible for refining and vertex recognition of the input regularized polygon sequence.

[0071] Vector output unit: generates final vectorized geographic target data based on vertex prediction results (for example, stored in formats such as Shapefile and GeoJSON).

[0072] Model training unit: Contains loss calculation and parameter optimization (such as gradient descent) functions, and is used to train the learnable modules in the system based on labeled data (the matching part of CAN and MCR may involve heuristic rules and PST).

[0073] These units can be implemented through software modules and work together to complete the overall extraction task from large-scale remote sensing images to vectorized geographic targets.

[0074] It should be noted that the specific algorithms mentioned in the above specific implementations (such as TC89-KCOS, Douglas-Peucker, UperNet, Transformer), network structure, loss function form, parameter settings (such as sampling rate, interpolation distance, etc.) , Tolerance , loss weight The above-mentioned examples are only given as examples. Those skilled in the art may adopt other algorithms with similar functions or make adaptive modifications based on actual application requirements and technological development. These modifications shall not depart from the scope of protection of the present invention.

[0075] In order to facilitate understanding of the technical effects of the present invention, a comparison between the present invention and the prior art is shown in Table 1: Table 1 Comparison of the invented method and the most advanced method on the WHU-building, GLH-water and VHR-road datasets

[0076] The method of the present invention excels in the vectorization of large-scale remote sensing images, surpassing existing state-of-the-art methods. Leveraging its unique attention mechanism, it can accurately extract large buildings and significantly outperforms other methods in metrics such as IoU, AP, and CIoU on the WHU-building dataset. Faced with the large-scale variations in water bodies in the GLH-water dataset, the method of the present invention demonstrates significant advantages, achieving significant improvements in IoU, AP, and CIoU, effectively addressing large-scale water targets. Despite the complex road network, the method of the present invention still performs best on the VHR-road dataset, generating roads with greater connectivity and less fragmentation. Its IoU, AP_l, and CIoU metrics all outperform other methods.

[0077] For simplicity of description, the method embodiments are described as a series of actions. However, those skilled in the art should be aware that the embodiments of the present invention are not limited by the order of the actions described, because certain steps can be performed in other orders or simultaneously according to the embodiments of the present invention. Secondly, those skilled in the art should also be aware that the embodiments described in this specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of the present invention.

[0078] In specific implementation, the method proposed in the technical solution of the present invention can be automatically run by those skilled in the art using computer software technology. System devices that implement the method, such as computer-readable storage media that store the corresponding computer program of the technical solution of the present invention and computer equipment that runs the corresponding computer program, should also be within the scope of protection of the present invention.

[0079] The electronic device provided by the present invention is described in the following embodiments. The electronic device described below and the method for extracting vector geographic targets from large-scale remote sensing images described above can be used for reference.

[0080] The electronic device may include a processor, a communications interface, a memory, and a communications bus, wherein the processor, the communications interface, and the memory communicate with each other via the communications bus. The processor may invoke logic instructions in the memory to execute a method for extracting vector geographic objects from large-scale remote sensing images, primarily including the software processing portion of the aforementioned steps.

[0081] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0082] On the other hand, an embodiment of the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the software processing part of the overall extraction method of vector geographic targets for large-format remote sensing images provided by the above methods.

[0083] On the other hand, an embodiment of the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the software processing part of the overall extraction method of vector geographic targets for large-scale remote sensing images provided by the above-mentioned methods.

[0084] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0085] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for extracting vector geographic targets from large-scale remote sensing images, characterized in that: include: In the context-aware segmentation stage, a contextual attention network is used to process the entire remote sensing image, extracting contextual information through a local-to-global attention mechanism to generate a segmentation mask for the target. In the mask edge reconstruction stage, the target outline is extracted from the segmentation mask and reconstructed through polygon simplification and fixed distance interpolation to generate a regularized polygon sequence with uniform vertex distribution; In the polygon sequence tracking stage, the regularized polygon sequence is input into the polygon sequence tracker, and the complete vectorized geographic target is output through position offset correction and vertex classification; Three of the stages perform end-to-end scene processing on large-scale remote sensing images without the need for segmentation.

2. The method for extracting vector geographic targets from large-scale remote sensing images according to claim 1, characterized in that: The working process of the context-aware segmentation stage includes the following processing: Construct a multi-scale pyramid of input images; Use sliding windows to synchronously extract image blocks of various scales; Extract features through an encoder with shared weights; The underlying high-resolution features are used as queries, multi-scale features are used as keys, and the attention mechanism is used to fuse contextual information. After concatenating the fusion features, the decoder outputs the segmentation mask.

3. The method for extracting vector geographic targets from large-scale remote sensing images according to claim 1, characterized in that: The working process of the mask edge reconstruction stage includes extracting the initial contour point set in the segmentation mask, simplifying the contour point set using a tolerance parameter, and performing linear interpolation at a fixed distance along each edge of the simplified polygon to generate a regularized polygon sequence.

4. The method for extracting vector geographic targets from large-scale remote sensing images according to claim 3 is characterized by: In the training of mask edge reconstruction, the reconstructed polygon sequence is matched with the true polygon, and vertex or non-vertex labels are assigned to the points in the sequence; and the true position data corresponding to the reconstructed sequence points are generated.

5. The method for extracting vector geographic targets from large-scale remote sensing images according to claim 1, characterized in that: When position offset correction is performed in the polygon sequence tracking stage, image features of each point in the regularized polygon sequence are extracted, and based on the image features and point coordinates, the position offset is predicted and the coordinates are corrected using a sequence model.

6. The method for extracting vector geographic targets from large-scale remote sensing images according to claim 5, characterized in that: When performing vertex classification in the polygon sequence tracking stage, the local geometric angle features of each point in the corrected polygon sequence are calculated, the coordinates, image features and angle features are integrated, the vertex probability is predicted, and the final vertex set is determined.

7. The method for extracting vector geographic targets from large-scale remote sensing images according to claim 6, characterized in that: A combined loss function is used when training the polygon sequence tracker, including the regression loss of position offset, the cross entropy loss of vertex classification, and the penalty loss based on geometric angles. The three types of losses are weighted and summed by preset weight coefficients.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory, wherein: When the processor executes the program, the method for overall extraction of vector geographic targets for large-scale remote sensing images as described in any one of claims 1 to 7 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for overall extraction of vector geographic targets from large-scale remote sensing images as described in any one of claims 1 to 7 is implemented.

10. A computer program product comprising a computer program, characterized in that: When the computer program is executed by a processor, the method for overall extraction of vector geographic targets from large-scale remote sensing images as described in any one of claims 1 to 7 is implemented.

Images