Deep learning-based method for recognition and extraction of ground features in 1: 2000 topographic map

Abstract

The present invention relates to the technical field of geographic information, and relates to a deep learning-based method for recognition and extraction of ground features in a 1: 2000 topographic map, comprising: in response to a delineated region input by a user terminal, obtaining a first ground feature symbol of a 1: 2000 topographic map; obtaining a first ground feature type matching result; when the first ground feature type matching result is null, obtaining a second ground feature type matching result; if the maximum size of a ground feature symbol size identifier is greater than or equal to 0.09 cm, performing size restoration to obtain a ground feature actual size; if the maximum size of the ground feature symbol size identifier is less than 0.09 cm, marking the ground feature actual size as null; and sending the second ground feature type matching result, the ground feature actual size, and a ground feature symbol position identifier to the user terminal. The present application solves the technical problem of insufficient accuracy and adaptability of existing topographic map ground feature recognition and extraction, and achieves the technical effect of improving the accuracy and adaptability of topographic map ground feature recognition and information extraction.

Description

A Deep Learning-Based Method for Feature Identification and Extraction from 1:2000 Topographic Maps Technical Field

[0001] This invention relates to the field of geographic information technology, and in particular to a method for identifying and extracting features from 1:2000 topographic maps based on deep learning. Background Technology

[0002] With the continuous development of Geographic Information Systems (GIS), unmanned topographic map feature recognition has been widely applied in fields such as traffic navigation and map culture display. In existing technologies, topographic map feature recognition typically employs a method based on map annotation symbols. This involves pre-setting a database linking annotation symbols to feature information, extracting map annotation symbols, and then combining this database to identify and extract information such as feature type, distribution, and size.

[0003] However, the existing technical solutions still have the following problems: First, topographic maps are diverse, and topographic maps from different sources and periods vary significantly in their representation, resulting in insufficient applicability of identification methods based on a single associated database. Second, even for annotation symbols representing the same feature, their size and shape may have subtle differences, making it difficult for such non-standardized symbols to be accurately matched and identified by the associated database. Third, limited by the cost and update efficiency of the associated database, existing solutions cannot cover all possible annotation symbols, and cannot achieve identification and information extraction when faced with unknown symbols. These problems result in insufficient accuracy and adaptability of existing solutions for topographic map feature identification and extraction, making it difficult to meet the practical application needs of unmanned topographic map feature identification. Summary of the Invention

[0004] This invention addresses the technical problems of insufficient accuracy and adaptability in existing topographic map feature identification and extraction methods due to the diversity of topographic map types, difficulties in recognizing non-standardized annotation symbols, and the inability to recognize unknown symbols. It provides a deep learning-based method for feature identification and extraction on 1:2000 topographic maps to solve these problems.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0006] This invention provides a method for feature identification and extraction on 1:2000 topographic maps based on deep learning, comprising: responding to a delineated area input by a user terminal, obtaining a first feature symbol on a 1:2000 topographic map, wherein the first feature symbol has a feature symbol size identifier and a feature symbol location identifier; processing the first feature symbol through a symbol description twin network and in conjunction with a symbol description database to obtain a first feature type matching result; when the first feature type matching result is empty, performing adjacency symbol analysis on the first feature symbol in conjunction with a topographic map sharing platform to obtain a second feature type matching result; when the maximum size of the feature symbol size identifier is greater than or equal to 0.09 cm, performing size restoration according to a 1:2000 scale to obtain the actual size of the feature; when the maximum size of the feature symbol size identifier is less than 0.09 cm, marking the actual size of the feature as empty; and sending the second feature type matching result, the actual size of the feature, and the feature symbol location identifier to the user terminal.

[0007] The beneficial effects of this invention are:

[0008] In response to the user-inputted designated area, the system obtains the first feature symbol from a 1:2000 topographic map. This first feature symbol includes both size and location identifiers. Map annotation symbols are retrieved from the user-designated area as objects to be identified, and their size and location information are extracted to provide input data for subsequent identification and information extraction. The first feature symbol is processed using a symbolic twin network combined with a symbolic database to obtain the first feature type matching result. The symbolic twin network is then introduced, and deep learning algorithms are used to learn and match the features of the annotation symbols. Combined with the symbolic database, preliminary identification of the feature type is achieved. When the first feature type matching result is empty, adjacency symbol analysis is performed on the first feature symbol using the topographic map sharing platform. The system obtains the matching results for the second feature type. For feature symbols that fail to match, it performs a second matching through adjacency symbol analysis to improve the recognition ability of non-standardized and unknown symbols. When the maximum size of the feature symbol size identifier is greater than or equal to 0.09 cm, it performs size restoration according to a 1:2000 scale to obtain the actual size of the feature. When the maximum size of the feature symbol size identifier is less than 0.09 cm, it marks the actual size of the feature as empty and performs size restoration calculation according to the map scale to obtain the actual size of the real feature. Features that are too small are filtered out to improve the accuracy of information extraction. The matching results of the second feature type, the actual size of the feature, and the location identifier of the feature symbol are sent to the user terminal, thereby improving the accuracy and applicability of topographic map feature recognition and extraction. Attached Figure Description

[0009] Figure 1 is a flowchart illustrating the deep learning-based method for identifying and extracting features from 1:2000 topographic maps provided by this invention.

[0010] Figure 2 is a schematic diagram of the electronic device provided by the present invention;

[0011] Figure 3 is a schematic diagram of the structure of a computer-readable storage medium provided by the present invention.

[0012] In the attached figures, the components represented by each number are as follows: electronic device 100, memory 110, processor 120, first computer program 111, computer-readable storage medium 200, and second computer program 211. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0015] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0016] Example 1:

[0017] As shown in Figure 1, this embodiment of the invention provides a method for feature identification and extraction from 1:2000 topographic maps based on deep learning, including:

[0018] S100: In response to the user's input of a defined area, obtain the first feature symbol of a 1:2000 topographic map, wherein the first feature symbol has a feature symbol size identifier and a feature symbol location identifier.

[0019] Specifically, users can access an image browsing interface containing the 1:2000 topographic map to be identified using a user-end device such as a smartphone, tablet, or PC. They can then manually or through other human-computer interaction methods to select the area of ​​interest on the image browsing interface. Subsequently, in response to the user's selected area, the image content within that area is used as the target data for subsequent identification. Next, feature symbols are sequentially obtained from the selected area, with each obtained feature symbol serving as the first feature symbol. A feature symbol is a symbolic representation of various land features (such as buildings, roads, and waterways) on a 1:2000 topographic map, using specific graphics. Different types of land features are typically represented by different symbols.

[0020] When extracting the first feature symbol, two associated attributes are also obtained: the feature symbol size identifier and the feature symbol location identifier. The feature symbol size identifier indicates the size of the symbol in the graphic, while the feature symbol location identifier indicates the spatial location of the symbol. These two attributes will be used in subsequent steps to reconstruct the actual size of the feature and determine its spatial location.

[0021] By responding to the user's delineation input, the first feature symbol representing the land feature entity is obtained from the target area of ​​the 1:2000 topographic map, and the size and location attribute information related to the symbol is extracted, which prepares for further identification of the land feature type represented by the symbol and acquisition of the real information of the land feature.

[0022] S200: The symbol of the first land feature is processed by the symbolic twin network and the symbolic database to obtain the matching result of the first land feature type.

[0023] Specifically, since different 1:2000 topographic maps may be produced by different cartographers, even when representing the same land cover type, the shape and color of the land cover symbols used may differ. To improve the accuracy and robustness of land cover symbol recognition, a symbol description twin network is used to process the first land cover symbol.

[0024] A Siamese network is a neural network structure that includes two subnetworks with identical structures and parameters. For example, a symbol description Siamese network might contain two parallel feature extraction channels and a post-processor feature comparison channel. The first feature symbol to be identified and a reference feature symbol randomly selected from the symbol description database are input into the two feature extraction channels, respectively, to extract their shape and color features. Then, the feature comparison channel calculates the similarity between the two sets of features. If both the shape and color similarity exceed a preset threshold, the category of the reference feature symbol (the first feature symbol) is taken as the identification result, i.e., the first feature type matching result.

[0025] The symbol description database is a collection of data storing the shape, color, and corresponding land cover category annotations of various standard land cover symbols. By continuously searching through the symbol examples in the database, the most similar symbol to the first land cover symbol can be found, thus revealing the land cover type represented by the first land cover symbol. If no match is found after traversing the symbol description database that satisfies both shape and color similarity requirements, the first land cover type matching result is set to empty, indicating that the semantic category of the first land cover symbol cannot be directly identified at this time.

[0026] By utilizing a symbolic twin network, the most similar reference symbols in shape and color are matched from the symbolic database to obtain the land cover type corresponding to the first land cover symbol, which is then used as the first land cover type matching result. This effectively overcomes the problem of differences in symbol representation across different topographic maps and improves the generalization ability of land cover symbol recognition.

[0027] S300: When the first feature type matching result is empty, the adjacency symbol analysis of the first feature symbol is performed in conjunction with the topographic map sharing platform to obtain the second feature type matching result.

[0028] Specifically, if the first feature type matching result is empty after processing with the symbol description twin network, it indicates that the feature category to which the feature symbol belongs cannot be determined solely by the symbol's shape and color. In this case, by combining the topographic map sharing platform, adjacency symbol analysis is performed on the first feature symbol to obtain the second feature type matching result.

[0029] Since 1:2000 topographic maps typically cover geomorphic information within a certain area, the types of land features represented by adjacent symbols on the map usually have some correlation. For example, rivers are often accompanied by bridges, and buildings are usually connected by roads. Therefore, the contextual semantic information formed by adjacent symbols can be used to indirectly infer the types of land features that cannot be directly identified at present.

[0030] To achieve the above objectives, a topographic map sharing platform is introduced. This platform is a cloud-based topographic map data sharing and service platform that stores a large amount of 1:2000 topographic map data drawn in different regions and at different times, and properly classifies, organizes, and semantically annotates this data. Through learning and analysis of massive amounts of topographic map data, the platform accumulates prior knowledge about the spatial relationships between different land cover types.

[0031] In practice, the first feature symbol is submitted to the topographic map sharing platform. The platform then searches for several adjacent symbols with similar shapes and colors. The feature type information of these adjacent symbols is summarized and statistically analyzed, and the feature type with the highest frequency is selected as the second feature type matching result. For example, if the symbol A to be identified is similar to symbols B, C, and D in the topographic map sharing platform, and the most frequent feature in the neighborhood of these three symbols is the "road" type, then it can be inferred that symbol A likely represents a road.

[0032] By making full use of the topographic map sharing platform, the semantic information of the symbols around the first feature symbol is analyzed and mined, and then a reasonable inference is made about the feature type of the symbol itself. This solves the problem that it is difficult to accurately identify the feature based on the symbol itself alone, and improves the overall accuracy of feature symbol identification.

[0033] S400: When the maximum size of the feature symbol is greater than or equal to 0.09 cm, the size is restored according to a 1:2000 scale to obtain the actual size of the feature.

[0034] Specifically, in 1:2000 topographic maps, feature symbols are typically drawn at a certain scaling ratio to ensure clear representation on the map. However, in practical applications, the actual size of features is often more important. Therefore, it is necessary to restore the symbol sizes to their actual dimensions according to the map's drawing scale.

[0035] At a scale of 1:2000, 1 centimeter on the map represents 20 meters in reality; that is, the ratio of map size to actual size is 1:2000. Based on practical experience, a threshold of 0.09 centimeters is used. When the maximum dimension of a feature symbol (the maximum of its length, width, and height) is greater than or equal to 0.09 centimeters, the actual feature size corresponding to that symbol is considered large enough to warrant reconstruction. When the maximum dimension is less than 0.09 centimeters, it can be ignored and no reconstruction is needed. This achieves a balance between the completeness of the reconstructed information and computational efficiency. For example, if the width of the first feature symbol to be processed is 0.13 centimeters and its height is 0.09 centimeters, its maximum dimension (0.13 centimeters) exceeds the threshold and needs to be reconstructed. Multiplying 0.13 by 2000 yields an actual width of 260 centimeters, or 2.6 meters. Since the height is equal to the threshold, it can also be reconstructed to 1.8 meters using the same method.

[0036] By multiplying ground feature symbols larger than a certain threshold by the reciprocal of the map scale, the actual size of the ground features can be calculated and restored. On the one hand, the true geometric information of the ground features contained in the symbols can be accurately obtained. On the other hand, by setting a threshold, minor errors that may exist under the limitations of map accuracy are eliminated, making the restoration results more reliable.

[0037] S500: When the maximum size of the feature symbol is less than 0.09cm, the actual size of the feature is marked as empty.

[0038] Specifically, as mentioned earlier, a threshold of 0.09 cm is set as the criterion for determining whether to restore the size of a feature symbol. When the maximum size of a feature symbol is less than this threshold, it can be considered that the actual feature corresponding to the symbol is either very small or difficult to distinguish from adjacent features. In this case, calculating and restoring its actual size is not very meaningful. Therefore, the actual size information of such feature symbols is directly marked as null and no restoration is performed.

[0039] For example, if the width of the feature symbol to be processed is 0.08 cm and the height is 0.06 cm, then its maximum size (0.08 cm) is less than the judgment threshold of 0.09 cm. In this case, the actual width and height of the feature symbol are marked as empty. The mark of empty can be represented by a specific character such as "null" or by a conventional value such as "-1", as long as it can be distinguished from the valid size value.

[0040] For land feature symbols that are too small, in order to ensure reliability and practicality, we do not restore their actual size, but only mark them as null values. This can effectively balance information quality and processing efficiency, and better support subsequent applications.

[0041] S600: Send the second feature type matching result, the actual size of the feature, and the location identifier of the feature symbol to the user terminal.

[0042] Specifically, the second feature type matching result represents the actual feature type represented by the first feature symbol, the actual size of the feature reflects its true geometric size, and the feature symbol location identifier indicates its spatial location. These three components constitute a comprehensive and accurate semantic description of a feature symbol, containing key information that users care about, such as feature type, size, and location.

[0043] The three extracted data points are packaged and transmitted in real-time over the network to the user client that initiated the recognition request. On the user client side, a specially designed information display interface presents the received semantic information of the ground features. For example, in the topographic map browsing interface, the text indicating the type of ground feature is overlaid or marked; the length, width, and height of the ground feature are displayed in pop-ups or dialog boxes; and the center point of the ground feature symbol is accurately marked in the map's spatial coordinate system. Users can intuitively and accurately grasp the semantic meaning of the original topographic map by combining the recognition results, significantly improving the readability and usability of the map information. Furthermore, in certain application scenarios, the user client can also convert the received semantic information of the ground features into other common data formats, such as XML and JSON, facilitating sharing and distribution across different systems and platforms.

[0044] By sending the results of feature symbol recognition and size restoration to the user's terminal via the network for display, the technical effect of improving the accuracy and applicability of topographic map feature recognition and information extraction is achieved.

[0045] Furthermore, embodiments of this application also include:

[0046] S110: Based on the defined area, the 1:2000 topographic map is segmented to obtain a segmented digital image;

[0047] S120: Using a multi-layer convolutional symbol detection model, perform semantic segmentation on the segmented digital image and extract a list of ground feature symbols, a list of ground feature symbol size identifiers, and a list of ground feature symbol location identifiers;

[0048] S130: The first feature symbol belongs to the feature symbol list, the feature symbol size identifier belongs to the feature symbol size identifier list, and the feature symbol location identifier belongs to the feature symbol location identifier list.

[0049] In one feasible implementation, the specific implementation of S100 can be further refined into steps S110 to S130, thereby illustrating how to obtain the first feature symbol and its related attributes from the 1:2000 topographic map based on the user's delineated area.

[0050] First, the region of interest (ROI) is identified by the user on their terminal device, either manually or through other human-computer interaction. Then, using this ROI as the cutting boundary, image segmentation is performed on the complete 1:2000 topographic map, resulting in a sub-image, called the segmented digital image. This segmented digital image contains only topographic map information within the ROI, reducing the data scale for subsequent recognition processing. Next, a multi-layer convolutional neural network (CNN) model based on deep learning is employed, specifically designed to detect and recognize ground feature symbols in the segmented digital image. Taking the segmented digital image as input, the CNN model accurately segments the graphic regions of each ground feature symbol from the complex image background through a series of operations such as convolution, pooling, and activation. Furthermore, the model can calculate the geometric dimensions (e.g., length, width, and height) of each symbol region, as well as its spatial location parameters in the image coordinate system. Finally, the model outputs three lists: a list of ground feature symbols, a list of ground feature symbol size identifiers, and a list of ground feature symbol location identifiers, storing the detected ground feature symbol graphics, their corresponding size information, and location information, respectively, preparing for subsequent processing.

[0051] Any feature symbol in the obtained list of feature symbols can be used as the first feature symbol for subsequent processing. This first feature symbol can be specified by the user, or it can be randomly selected by the system or selected according to certain rules. At the same time, the feature symbol size identifier of the first feature symbol can be extracted from the feature symbol size identifier list, and the feature symbol location identifier of the first feature symbol can be extracted from the feature symbol location identifier list.

[0052] First, the original topographic map is segmented according to the user-defined area. Then, a specially designed multi-layer convolutional symbol detection model is used to identify ground feature symbols in the segmented sub-maps, extracting the graphic, size, and location information of the ground feature symbols. Any symbol is then identified as the target symbol for subsequent processing, providing a foundation for subsequent ground feature identification and extraction.

[0053] Furthermore, embodiments of this application also include:

[0054] S121: Set the symbol detection size threshold for the 1:2000 topographic map, wherein the symbol detection size threshold is less than or equal to 0.33 mm;

[0055] S122: Initialize the feature pyramid topology according to the symbol detection size threshold to obtain several initial feature pyramid topologies;

[0056] S123: Train the aforementioned initial feature pyramid topologies to obtain several feature symbol segmentation accuracies, wherein the feature symbol segmentation accuracy is represented by the similarity between the feature segmentation symbol and the preset supervision symbol of the feature.

[0057] S124: When the segmentation accuracy of the plurality of ground feature symbols is less than or equal to the convergence accuracy threshold, the feature pyramid topology is optimized based on the segmentation accuracy of the plurality of ground feature symbols to obtain a feature pyramid target topology that is greater than the convergence accuracy threshold.

[0058] S125: When the segmentation accuracy of the plurality of ground feature symbols has an initial feature pyramid topology that is greater than the convergence accuracy threshold, it is set as the target feature pyramid topology;

[0059] S126: Generate the multi-layer convolutional symbol detection model based on the target topology of the feature pyramid.

[0060] In a preferred embodiment, detailed steps for constructing a multi-layer convolutional symbol detection model are provided, namely steps S121 to S126, to illustrate how to design and optimize the network structure and parameters of the detection model for a 1:2000 topographic map to achieve high-precision and robust symbol recognition.

[0061] First, a symbol detection size threshold is set for 1:2000 topographic maps, where the threshold is less than or equal to 0.33 mm. According to relevant standards, the minimum size requirement for feature symbols in 1:2000 topographic maps is 0.33 mm. Based on this, a symbol detection size threshold specifically for this scale of topographic map is set to provide prior knowledge for subsequent network training. Then, the feature pyramid topology is initialized according to the symbol detection size threshold, obtaining several initial feature pyramid topologies. A feature pyramid is a multi-scale, hierarchical convolutional neural network structure capable of extracting image features at different receptive fields and resolutions. Using the symbol detection size threshold as a constraint, a set of initial feature pyramid network topologies is randomly generated as the starting point for subsequent structure search. Subsequently, several initial feature pyramid topologies are trained to obtain several feature symbol segmentation accuracies, where the feature symbol segmentation accuracy is represented by the similarity between the segmented feature symbol and the preset supervised feature symbol. Specifically, each initial topology is trained using a manually annotated topographic map dataset, and its symbol segmentation accuracy is evaluated on a validation set. The segmentation accuracy is defined based on the similarity measure between the segmented symbols and the manually labeled supervised symbols.

[0062] When the segmentation accuracy of several ground feature symbols is less than or equal to the convergence accuracy threshold, feature pyramid topology optimization is performed on several initial feature pyramid topologies based on their segmentation accuracy to obtain a target feature pyramid topology with a convergence accuracy greater than the threshold. Specifically, if the segmentation accuracy of existing topologies is not ideal, hyperparameters such as the number of network layers, filter size, and number of filters for each topology are iteratively optimized until the accuracy of a certain structure exceeds a preset convergence threshold. This structure is then determined as the target feature pyramid topology to adapt to the characteristics of the current data and task. When several ground feature symbol segmentation accuracies have an initial feature pyramid topology with a convergence accuracy greater than the threshold, it is set as the target feature pyramid topology. Specifically, if the segmentation accuracy of an initial topology has already reached the convergence threshold, no structural optimization is needed, and it is directly determined as the target feature pyramid topology to save computational costs. Subsequently, the obtained target feature pyramid topology is used to generate a corresponding multi-layer convolutional neural network model to complete the construction of the detection model.

[0063] By initializing, training, evaluating, and dynamically optimizing the feature pyramid topology under a specially designed size threshold constraint, the optimal network structure is automatically searched, thereby generating a customized detection model. While ensuring detection accuracy, the model design process is automated to the greatest extent and the cost of manual trial and error is reduced.

[0064] Furthermore, embodiments of this application also include:

[0065] S1241: Construct a feature pyramid topological distance evaluation function, wherein the feature pyramid topological distance evaluation function is used to evaluate the distance feature value between any two feature pyramid topologies;

[0066] S1242: Based on the feature pyramid topological distance evaluation function and a topological distance threshold, perform cluster analysis on the several initial feature pyramid topologies to obtain multi-cluster initial feature pyramid topologies;

[0067] S1243: Based on the segmentation accuracy of the aforementioned feature symbols, extract the initial feature pyramid topology with the highest accuracy of the multi-cluster initial feature pyramid topology to obtain multiple initial feature pyramid topologies with the highest accuracy.

[0068] S1244: Set the multiple initial feature pyramid topologies with the highest accuracy as multiple guiding feature pyramid topologies, perform guiding mutation on the non-cluster initial feature pyramid topologies, and obtain an updated feature pyramid topology, wherein guiding mutation refers to shortening the distance between the guided feature pyramid topology and the guiding feature pyramid topology.

[0069] S1245: Train the updated feature pyramid topology to obtain the updated feature symbol segmentation accuracy;

[0070] S1246: When the updated feature symbol segmentation accuracy is greater than the convergence accuracy threshold, the updated feature pyramid topology is set as the target feature pyramid topology; otherwise, the feature pyramid topology optimization loop is executed.

[0071] In a preferred embodiment, the specific implementation of optimizing several initial feature pyramid topologies can be further refined into steps S1241 to S1246. By introducing mechanisms such as feature pyramid topology distance evaluation function, cluster analysis, and guided mutation, the optimal network structure can be searched more intelligently and efficiently.

[0072] First, a feature pyramid topological distance evaluation function is constructed to quantitatively characterize the distance feature value between any two topologies, representing the degree of difference between them. This function considers factors such as the number of layers in the topology, the number of filter types in each layer, and the receptive field size, calculating the comprehensive difference value through weighted summation, providing a distance metric for subsequent clustering analysis. Next, using the constructed feature pyramid topological distance evaluation function, several initial feature pyramid topologies are clustered with a topological distance threshold as the radius, grouping structurally similar topologies into the same cluster, resulting in multi-cluster initial feature pyramid topologies. Multi-cluster initial feature pyramid topologies represent coarse-grained grouping of the initial topologies, with topologies within each cluster sharing locally optimal substructures, thus helping to narrow the scope of subsequent optimization searches and improve optimization efficiency. Subsequently, based on the segmentation accuracy of several land cover symbols, the topology with the highest segmentation accuracy is selected from the initial feature pyramid topologies of each cluster as the highest-accuracy initial feature pyramid topology in the corresponding cluster, resulting in multiple highest-accuracy initial feature pyramid topologies, representing high-quality topologies within each cluster, which can guide the optimization direction of other topologies.

[0073] Then, multiple initial feature pyramid topologies with the highest accuracy are used as multiple guiding feature pyramid topologies. Using these guiding feature pyramid topologies as references, the structural parameters of non-cluster initial feature pyramid topologies are adjusted so that they gradually approach the guiding feature pyramid topologies under the definition of the feature pyramid topology distance evaluation function, completing iterative optimization of the structure. The step size of the guiding mutation can be dynamically controlled to balance optimization speed and exploration breadth. After guiding mutation, a set of updated feature pyramid topologies is generated. Then, the feature symbol segmentation accuracy of the updated feature pyramid topologies is re-evaluated. Once the accuracy of an updated topology exceeds a preset convergence accuracy threshold, it is determined as the final target feature pyramid topology, and the optimization process ends; otherwise, clustering, guiding mutation, and other steps continue iteratively until the convergence condition is met or the maximum number of iterations is reached.

[0074] By integrating mechanisms such as distance metrics, cluster analysis, and guided mutation, an effective feature pyramid topology optimization strategy is formed. This strategy can quickly lock the optimal network structure in the search space while also taking into account the optimization of local details, providing a strong guarantee for the high-precision and adaptive construction of multi-layer convolutional symbol detection models.

[0075] Furthermore, the feature pyramid topological distance evaluation function is: X=[(x 11 ,x 12 ),…,(x i1 ,x i2 ),…,(x N1 ,x N2 )] Y=[(y 11 ,y 12 ),…,(y i1 ,y i2 ),…,(y M1 ,y M2 )]

[0076] Where D(X,Y) represents the distance between the two feature pyramid topologies, X represents the first feature pyramid topology, Y represents the second feature pyramid topology, and x... 11 x 12 The dimension of the first-layer convolutional kernel in both directions, x, representing the topology of the first feature pyramid. i1 x i2 The dimension of the i-th layer convolutional kernel in two directions, representing the topology of the first feature pyramid. N1 x N2 The dimension of the Nth layer convolutional kernel representing the topology of the first feature pyramid is defined by the dimensions of the two directions, where N represents the total number of layers in the first feature pyramid topology. 11 y 12 The dimension of the first convolutional kernel representing the topology of the second feature pyramid, y, in both directions. i1 y i2 The dimensions of the i-th layer convolution kernel representing the topology of the second feature pyramid, y, are in both directions. M1 y M2 The dimension of the M-th layer convolution kernel representing the second feature pyramid topology is defined by the two directions, where M represents the total number of layers in the second feature pyramid topology.

[0077] In a preferred embodiment, the constructed feature pyramid topological distance evaluation function can take the following form: X=[(x 11 ,x 12 ),…,(x i1 ,x i2 ),…,(x N1 ,x N2 )] Y=[(y 11 ,y12 ),…,(y i1 ,y i2 ),…,(y M1 ,y M2 )]

[0078] First, define two feature pyramid topologies X and Y to be compared, where X consists of N convolutional kernels and Y consists of M convolutional kernels. The two dimensions of the i-th convolutional kernel in X are denoted by x. i1 and x i2 This means that the two dimensions of the i-th convolutional kernel in Y are represented by y. i1 and y i2 This is represented as follows. Then, the distance D(X,Y) between the feature pyramid topologies X and Y is defined as:

[0079] Where f(M,N) is a binary function, that is:

[0080] The value is MN when M is not equal to N, and 1 when M is equal to N. min(M,N) means taking the smaller value between M and N.

[0081] The core idea of ​​the aforementioned feature pyramid topological distance evaluation function is as follows: First, f(M,N) is used to measure the difference in the number of layers between two feature pyramid topologies; the greater the difference in the number of layers, the greater the distance between the two topologies. Second, for two topologies with the same number of layers, the Euclidean distance of the difference in their convolutional kernel dimensions is calculated layer by layer, and the results are summed to obtain the total distance metric. Therefore, the difference in the number of layers determines the order of magnitude of the topological distance, while the difference in the convolutional kernel dimensions determines the fine-grained ranking of topologies with the same number of layers.

[0082] The aforementioned feature pyramid topological distance evaluation function is only one possible implementation, and its specific form can be adjusted according to actual needs. However, regardless of the form adopted, the basic starting point is the same: to comprehensively consider the number of layers and the convolution kernel dimension of the feature pyramid topology, quantitatively characterize the degree of difference of the feature pyramid topology in the structural feature space, and provide an operable distance metric for cluster analysis and guided mutation.

[0083] By providing a feature pyramid topological distance evaluation function, the hierarchical structure information of the feature pyramid topology is fully utilized, and the topological comparison is transformed into a comprehensive measure of the number of layers and the dimension of the convolution kernel, laying the foundation for building a multi-layer convolutional symbol detection model.

[0084] Furthermore, the notation indicates that the Siamese network includes a first feature extraction channel, a second feature extraction channel, and a feature comparison channel. The first feature extraction channel and the second feature extraction channel have the same structural topology and parameters. The first land cover type matching result is obtained, including:

[0085] S210: Input the first feature symbol into the first feature extraction channel to obtain the shape feature and color feature of the first feature symbol;

[0086] S220: Input the first reference feature symbol randomly extracted from the symbol description database into the second feature extraction channel to obtain the shape feature and color feature of the second feature symbol;

[0087] S230: Input the shape features of the first feature symbol, the color features of the first feature symbol, the shape features of the second feature symbol, and the color features of the second feature symbol into the feature comparison channel to obtain shape similarity and color similarity;

[0088] S240: When the shape similarity is greater than or equal to the shape similarity threshold and the color similarity is greater than or equal to the color similarity threshold, the associated land feature type of the first reference land feature symbol is set as the first land feature type matching result; otherwise, the first reference land feature symbol is updated for cyclic matching according to the symbol description database.

[0089] S250: When the symbol indicates that there is no matching result for the first land feature type in the database, the matching result for the first land feature type is set to empty.

[0090] In a preferred embodiment, steps S210 to S250 are performed to describe in detail the internal structure and working principle of the twin network using symbols.

[0091] First, the symbol description Siamese network consists of three sub-networks: a first feature extraction channel, a second feature extraction channel, and a feature comparison channel. The first and second feature extraction channels use identical topologies and parameters. Then, the first feature symbol to be identified is input into the first feature extraction channel. After a series of convolution and pooling operations, the shape and color features of the symbol are extracted, serving as the first feature symbol's shape and color features. The shape features describe the symbol's geometric appearance, while the color features reflect its color attributes. Simultaneously, a reference feature symbol, called the first reference feature symbol, is selected from the symbol description database and input into the second feature extraction channel. Similarly, the second feature extraction channel outputs the shape and color features of this reference symbol, serving as the second feature symbol's shape and color features.

[0092] Subsequently, the shape and color features of the first feature symbol and the first reference feature symbol are summarized, namely, the shape features of the first feature symbol, the color features of the first feature symbol, the shape features of the second feature symbol, and the color features of the second feature symbol, and input into the feature comparison channel. The feature comparison channel uses a similarity metric to calculate the similarity in shape between the shape features of the first feature symbol and the shape features of the second feature symbol, and also calculates the similarity in shape between the color features of the first feature symbol and the color features of the second feature symbol, obtaining two indices: shape similarity and color similarity. Then, a threshold judgment is applied to the shape similarity and color similarity. If both shape similarity and color similarity exceed a preset threshold, i.e., shape similarity is greater than or equal to the shape similarity threshold, and color similarity is greater than or equal to the color similarity threshold, then the first feature symbol and the first reference feature symbol are considered to belong to the same category. In this case, the associated feature type noted in the symbol description database of the first reference feature symbol is output as the first feature type matching result. Otherwise, if the current first reference feature symbol is deemed insufficiently similar, it needs to be updated, and steps S220 to S240 are repeated until a sufficiently similar matching symbol is found, or until the entire symbol description database is traversed. If no reference feature symbol matching the first feature symbol exists in the symbol description database, meaning no valid first feature type match is ultimately obtained, then the value is set to null, indicating that there is no category matching the first feature symbol in the symbol description database.

[0093] By using symbolic twin networks for type identification of ground feature symbols, we can effectively address the diversity and complexity of ground feature symbols and significantly improve the accuracy and efficiency of symbol identification.

[0094] Furthermore, implementation of this application also includes:

[0095] S310: Set the shape similarity adjacency threshold and the color similarity adjacency threshold;

[0096] S320: Based on the shape similarity adjacency threshold and the color similarity adjacency threshold, match several adjacent land cover types for the first land cover symbol from the topographic map sharing platform;

[0097] S330: Extract the highest frequency land cover type from the plurality of land cover types and set it as the second land cover type matching result.

[0098] In a preferred embodiment, steps S310 to S330 further explain how to use the topographic map sharing platform to perform further adjacency symbol analysis on the first feature symbol to obtain the second feature type matching result when the first feature type matching result is empty.

[0099] First, two threshold parameters are set: a shape similarity adjacency threshold and a color similarity adjacency threshold. These two thresholds are used to retrieve adjacent symbols similar to the first feature symbol in the topographic map sharing platform, but their values ​​are usually higher than the shape similarity threshold and color similarity threshold. This is because, in the absence of direct matches, appropriately relaxing the similarity requirements of adjacent symbols helps to obtain more reference information. Subsequently, based on the above two adjacency thresholds, adjacent feature symbols similar in shape and color to the first feature symbol are retrieved in the topographic map sharing platform. The retrieved adjacent feature symbols and their corresponding feature types are provided as candidate matches for the next step. Then, the frequency of each feature type appearing in all adjacent feature symbols is counted, and the feature type with the highest frequency is determined as the second feature type match result. This utilizes the prior assumption that spatially adjacent and feature-similar feature symbols tend to belong to a consistent feature type. Frequency statistics can, to some extent, eliminate the interference of individual abnormal matches and improve the confidence of the matching results.

[0100] When direct matching fails, the system leverages the massive amount of feature symbol data and implicit spatial topological relationships contained in the topographic map sharing platform to infer the type of the first feature symbol. This adjacency analysis-based inference mechanism compensates for the limitations of single symbol shape and color features, enabling the discovery of matching clues over a wider range and improving the success rate of identification.

[0101] Furthermore, embodiments of this application also include:

[0102] S321: Based on the shape similarity adjacency threshold and the color similarity adjacency threshold, match first-level adjacent feature symbols for the first feature symbol from the topographic map sharing platform;

[0103] S322: When the data volume of the first-level adjacent object symbol is less than or equal to the data volume threshold, based on the shape similarity adjacency threshold and the color similarity adjacency threshold, the first-level adjacent object symbols are traversed from the topographic map sharing platform to match the second-level adjacent object symbols;

[0104] S323: Add the secondary adjacent object symbol and the primary adjacent object symbol to the plurality of adjacent object symbols.

[0105] In a preferred embodiment, the matching range of adjacency symbols is expanded by introducing first-level adjacency and second-level adjacency in a hierarchical and progressive manner.

[0106] First, based on shape similarity and color similarity adjacency thresholds, topographic map sharing platforms are used to retrieve topographic map symbols that meet these thresholds; these are called first-level adjacency symbols. Then, based on the first-level adjacency matching, second-level adjacency is introduced. When the number of first-level adjacency symbols is small, i.e., less than a preset data threshold, the topographic map sharing platform is further searched for topographic map symbols that meet the shape and color similarity thresholds, based on the first-level adjacency symbols, to obtain second-level adjacency symbols. This expands the number of topographic map symbols to obtain richer information when the number of first-level adjacency symbols is small. Subsequently, the obtained first-level and second-level adjacency symbols are merged as several adjacency symbols of the first topographic map symbol, preparing for subsequent frequency statistical analysis.

[0107] By adopting a hierarchical adjacency symbol matching strategy that proceeds from near to far and from direct to indirect, the robustness and adaptability to complex feature symbols are enhanced.

[0108] The method for feature identification and extraction based on deep learning in 1:2000 topographic maps provided in this invention has at least the following technical effects:

[0109] Responding to the user-inputted delineated area, the system obtains the first feature symbol from a 1:2000 topographic map. This first feature symbol includes both size and location identifiers, providing a foundation for subsequent identification and information extraction. A symbol description twin network, combined with a symbol description database, is used to process the first feature symbol, obtaining a first feature type matching result, thus improving the recognition capability of complex, non-standardized symbols. When the first feature type matching result is empty, adjacency symbol analysis is performed on the first feature symbol using a topographic map sharing platform to obtain a second feature type matching result, further improving the recall rate and compensating for the limitations of single-feature matching. When the maximum size of the feature symbol's size identifier is greater than or equal to 0.09 cm, the actual size of the feature is restored according to the 1:2000 scale to obtain the actual size. When the maximum size of the feature symbol's size identifier is less than 0.09 cm, the actual size is marked as empty, extracting valuable size information and thus improving the accuracy of the information extraction results. The matching results of the second feature type, the actual size of the feature, and the location identifier of the feature symbol are sent to the user terminal, thereby improving the accuracy and adaptability of topographic map feature identification and information extraction.

[0110] Example 2:

[0111] Please refer to Figure 2, which is a schematic diagram of an embodiment of the electronic device provided in this invention. As shown in Figure 2, an electronic device 100 provided in this invention includes a memory 110, a processor 120, and a first computer program 111 stored in the memory 110 and executable on the processor 120. When the processor 120 executes the first computer program 111, it implements a method for identifying and extracting features from a 1:2000 topographic map based on deep learning.

[0112] Example 3:

[0113] Please refer to Figure 3, which is a schematic diagram of an embodiment of a computer-readable storage medium provided by an embodiment of the present invention. As shown in Figure 3, this embodiment provides a computer-readable storage medium 200, on which a second computer program 211 is stored. When the second computer program 211 is executed by a processor, it implements a method for identifying and extracting features from a 1:2000 topographic map based on deep learning.

[0114] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0115] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0116] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or one or more blocks of the block diagrams.

[0117] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowcharts and / or one or more block diagrams.

[0118] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0119] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0120] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for feature identification and extraction from 1:2000 topographic maps based on deep learning, characterized in that, include: In response to the delineated area input by the user, a first feature symbol of a 1:2000 topographic map is obtained, wherein the first feature symbol has a feature symbol size identifier and a feature symbol location identifier; By using a symbolic twin network and combining it with a symbolic database, the first land cover symbol is processed to obtain the first land cover type matching result; When the matching result of the first land feature type is empty, the adjacency symbol analysis of the first land feature symbol is performed in conjunction with the topographic map sharing platform to obtain the matching result of the second land feature type. When the maximum size of the feature symbol is greater than or equal to 0.09 cm, the actual size of the feature is obtained by restoring the size according to a 1:2000 scale. When the maximum size of the feature symbol is less than 0.09 cm, the actual size of the feature is marked as empty; The matching result of the second feature type, the actual size of the feature, and the location identifier of the feature symbol are sent to the user terminal.

2. The method as described in claim 1, characterized in that, In response to the user's input of a defined area, the first feature symbol of the 1:2000 topographic map is obtained, including: Based on the defined area, the 1:2000 topographic map is segmented to obtain a segmented digital image; The segmented digital image is semantically segmented using a multi-layer convolutional symbol detection model to extract a list of ground feature symbols, a list of ground feature symbol size identifiers, and a list of ground feature symbol location identifiers. The first feature symbol belongs to the feature symbol list, the feature symbol size identifier belongs to the feature symbol size identifier list, and the feature symbol location identifier belongs to the feature symbol location identifier list.

3. The method as described in claim 2, characterized in that, The steps for constructing the multi-layer convolutional sign detection model include: Set a symbol detection size threshold for the 1:2000 topographic map, wherein the symbol detection size threshold is less than or equal to 0.33 mm; Based on the symbol detection size threshold, the feature pyramid topology is initialized to obtain several initial feature pyramid topologies; The initial feature pyramid topology is trained to obtain several feature symbol segmentation accuracies, wherein the feature symbol segmentation accuracy is represented by the similarity between the feature segmentation symbol and the preset supervision symbol of the feature. When the segmentation accuracy of the plurality of ground feature symbols is less than or equal to the convergence accuracy threshold, the feature pyramid topology is optimized based on the segmentation accuracy of the plurality of ground feature symbols to obtain a feature pyramid target topology that is greater than the convergence accuracy threshold. When the segmentation accuracy of the plurality of ground feature symbols has an initial feature pyramid topology that is greater than the convergence accuracy threshold, it is set as the target feature pyramid topology; The multi-layer convolutional symbol detection model is generated based on the target topology of the feature pyramid.

4. The method as described in claim 3, characterized in that, Based on the segmentation accuracy of the aforementioned feature symbols, feature pyramid topology optimization is performed on the aforementioned initial feature pyramid topologies to obtain a target feature pyramid topology with a convergence accuracy threshold, including: Construct a feature pyramid topological distance evaluation function, wherein the feature pyramid topological distance evaluation function is used to evaluate the distance feature value between any two feature pyramid topologies; Based on the feature pyramid topological distance evaluation function, and based on the topological distance threshold, cluster analysis is performed on the several initial feature pyramid topologies to obtain multi-cluster initial feature pyramid topologies; Based on the segmentation accuracy of the aforementioned feature symbols, the highest accuracy initial feature pyramid topology of the multi-cluster initial feature pyramid topology is extracted to obtain multiple highest accuracy initial feature pyramid topologies. The multiple initial feature pyramid topologies with the highest accuracy are set as multiple guiding feature pyramid topologies. Guiding mutations are performed on the initial feature pyramid topologies that are not in this cluster to obtain updated feature pyramid topologies. Here, guiding mutation refers to shortening the distance between the guided feature pyramid topology and the guiding feature pyramid topology. The updated feature pyramid topology is trained to obtain the updated feature symbol segmentation accuracy. If the updated feature symbol segmentation accuracy is greater than the convergence accuracy threshold, the updated feature pyramid topology is set as the target feature pyramid topology; otherwise, a feature pyramid topology optimization loop is executed.

5. The method as described in claim 4, characterized in that, The topological distance evaluation function of the feature pyramid is: X=[(x 11 ,x 12 ),…,(x i1 ,x i2 ),…,(x N1 ,x N2 )] Y=[(y 11 ,y 12 ),…,(y i1 ,y i2 ),…,(y M1 ,y M2 )] Where D(X,Y) represents the distance between the two feature pyramid topologies, X represents the first feature pyramid topology, Y represents the second feature pyramid topology, and x... 11 x 12 The dimension of the first-layer convolutional kernel in both directions, x, representing the topology of the first feature pyramid. i1 x i2 The dimension of the i-th layer convolutional kernel in two directions, representing the topology of the first feature pyramid. N1 x N2 The dimension of the Nth layer convolutional kernel representing the topology of the first feature pyramid is defined by the dimensions of the two directions, where N represents the total number of layers in the first feature pyramid topology. 11 y 12 The dimension of the first convolutional kernel representing the topology of the second feature pyramid, y, in both directions. i1 y i2 The dimensions of the i-th layer convolution kernel representing the topology of the second feature pyramid, y, are in both directions. M1 y M2 The dimension of the M-th layer convolution kernel representing the second feature pyramid topology is defined by the two directions, where M represents the total number of layers in the second feature pyramid topology.

6. The method as described in claim 1, characterized in that, The symbolic twin network includes a first feature extraction channel, a second feature extraction channel, and a feature comparison channel. The first feature extraction channel and the second feature extraction channel have the same structural topology and parameters. The symbolic twin network, combined with a symbolic database, processes the first land cover symbol to obtain a first land cover type matching result, including: Input the first feature symbol into the first feature extraction channel to obtain the shape feature and color feature of the first feature symbol; The first reference feature symbol randomly extracted from the symbol description database is input into the second feature extraction channel to obtain the shape feature and color feature of the second feature symbol. The shape features of the first feature symbol, the color features of the first feature symbol, the shape features of the second feature symbol, and the color features of the second feature symbol are input into the feature comparison channel to obtain shape similarity and color similarity; When the shape similarity is greater than or equal to the shape similarity threshold and the color similarity is greater than or equal to the color similarity threshold, the associated land feature type of the first reference land feature symbol is set as the first land feature type matching result; otherwise, the first reference land feature symbol is updated for cyclic matching according to the symbol description database. When the symbol indicates that there is no matching result for the first land cover type in the database, the matching result for the first land cover type is set to empty.

7. The method as described in claim 1, characterized in that, When the first feature type matching result is empty, adjacency symbol analysis is performed on the first feature symbol using the topographic map sharing platform to obtain the second feature type matching result, including: Set adjacency thresholds for shape similarity and color similarity; Based on the shape similarity adjacency threshold and the color similarity adjacency threshold, from the topographic map sharing platform, several adjacent land cover types of the first land cover symbol are matched. Extract the highest frequency land cover type from the given land cover types and set it as the second land cover type matching result.

8. The method as described in claim 7, characterized in that, Based on the shape similarity adjacency threshold and the color similarity adjacency threshold, from the topographic map sharing platform, several adjacent feature symbols and several feature types are matched for the first feature symbol, including: Based on the shape similarity adjacency threshold and the color similarity adjacency threshold, first-level adjacent feature symbols are matched for the first feature symbol from the topographic map sharing platform; When the data volume of the first-level adjacent object symbol is less than or equal to the data volume threshold, based on the shape similarity adjacency threshold and the color similarity adjacency threshold, the first-level adjacent object symbols are traversed from the topographic map sharing platform to match the second-level adjacent object symbols; The secondary adjacent object symbols and the primary adjacent object symbols are added to the plurality of adjacent object symbols.

Images