Method and system for identifying midpoint symbols in scanned topographic map by using deep convolutional neural network

Abstract

The invention relates to the technical field of computer map making, in particular to a method and system for identifying point symbols in a scanned topographic map by using a deep convolutional neural network, and the method comprises the steps: constructing a point symbol data set suitable for supervised learning, and dividing the data set in proportion; expanding the data set of the original map in the data set by using a data enhancement method; based on a k-means + + clustering algorithm, generating an anchor frame suitable for multiple types of point symbols; designing an ASP-YOLOv4 model, and carrying out the positioning and detection of a plurality of types of point symbols; the whole topographic map is subjected to image processing operation of firstly cutting the map, then amplifying, then detecting and finally splicing; and training and evaluating the model by using the loss function and the data set. According to the method, the cavity space pyramid pooling module is introduced, so that multi-scale information is effectively obtained, and the recognition precision and the positioning accuracy of the point symbols are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of computer map making, and in particular to a method and system for identifying point symbols in a scanned topographic map using a deep convolutional neural network. Background Art

[0002] Scanned topographic maps are widely used in resource development, engineering construction, and military defense. The point symbols in them provide important geographic information. However, the recognition of point symbols faces multiple challenges: point symbols are small in size, similar in shape, and often overlap with other geographic elements (such as contour lines). Traditional image segmentation and pattern recognition methods have difficulty effectively distinguishing these symbols. In addition, when the scanning quality is not high, the point symbols may be blurred, deformed, or broken, further increasing the difficulty of recognition.

[0003] To overcome these challenges, a variety of methods have been proposed for the recognition of point symbols in scanned topographic maps, mainly including template matching, statistical decision making, grammatical structure and neural network methods. The earliest point symbol recognition method is based on template matching. Although this method is easy to implement, it is easily interfered by background noise, resulting in poor recognition performance. To improve this problem, researchers have proposed improved template matching methods. For example, fuzzy classification and recognition methods are used to identify oil well identifiers in scanned petroleum geological maps. In complex scenarios, the shear segment generalized Hough transform (SLS-GHT) method based on the Hough transform algorithm was proposed (Miao et al., 2017) for point symbol recognition in scanned topographic maps, with good recognition results. However, the SLS-GHT method requires the construction of an R table, which is complex and time-consuming, and has poor generalization ability. In addition, Reiher et al. (1996) used Hausdorff distance and neural networks to recognize point symbols; Leyk and Chiang (2006) first introduced the concept of geographic context to automatically recognize symbols in scanned maps; Pezeshk et al. (2011) used a series of morphological operations for symbol recognition.

[0004] With the development of computer vision, deep convolutional neural networks (DCNNs) have been proposed and widely used in image processing tasks such as target detection and semantic segmentation. Deep learning technology has achieved remarkable results in the fields of remote sensing image target detection and ground object change recognition, and has made progress in symbol recognition tasks in multiple fields. For example, Quan et al. (2018) proposed a symbol classifier architecture based on deep transfer learning, using AlexNet to classify and recognize multi-point symbols to improve recognition efficiency. In geological map symbol extraction and recognition, the method combining DCNN and graph convolutional network (GCN) has also achieved good results (Guo et al., 2021). However, there are significant differences between the point symbols in scanned topographic maps and those in traditional images. Their shapes are more abstract and closely integrated with the background. Existing symbol recognition methods have failed to effectively solve these problems. Summary of the invention

[0005] The present invention aims to solve the problem that point symbols in scanned topographic maps provide key geographic information, but due to the adhesion between topographic map elements and the singleness of symbol structure, the recognition of point symbols faces high complexity and uncertainty, resulting in low accuracy and efficiency of traditional symbol recognition methods. A method and system for point symbol recognition in scanned topographic maps using a deep convolutional neural network are proposed, and an atrous spatial pyramid pooling (ASPP) module is introduced to effectively obtain multi-scale information, thereby significantly improving the recognition accuracy and positioning accuracy of point symbols.

[0006] To achieve the above purpose, the technical solution adopted is:

[0007] The present invention provides a method for identifying point symbols in a scanned topographic map using a deep convolutional neural network, comprising:

[0008] Construct a point symbol dataset suitable for supervised learning and divide the dataset proportionally;

[0009] Use data augmentation methods to expand the dataset for the original maps in the dataset;

[0010] Generate anchor boxes suitable for multiple types of point symbols based on the k-means++ clustering algorithm;

[0011] Design the ASPP-YOLOv4 model to locate and detect multiple types of point symbols;

[0012] The entire topographic map is processed by the image processing operation of "first cutting, then enlarging, then detecting, and finally splicing";

[0013] Use loss functions and datasets to train and evaluate the model.

[0014] According to the present invention, a method for identifying point symbols in a scanned topographic map is performed using a deep convolutional neural network. Further, a point symbol data set constructed includes scanned topographic maps of a fixed scale; point symbols are annotated for each scanned topographic map in the data set, and the category of the symbol and its coordinates in the image are recorded.

[0015] According to the present invention, a method for identifying point symbols in a scanned topographic map using a deep convolutional neural network is used. Further, the data enhancement method includes horizontal flipping, vertical flipping, random cropping, graying, Gaussian blurring, color jittering and small-angle rotation of the original image.

[0016] According to the present invention, a method for identifying point symbols in a scanned topographic map using a deep convolutional neural network is further used to generate anchor frames suitable for multiple types of point symbols based on a k-means++ clustering algorithm, including:

[0017] The k-means++ clustering algorithm is used to obtain multiple clusters through iterative calculation according to the anchor box size of the symbols in the data set; the algorithm selects the optimal anchor box size for each cluster by calculating the intersection ratio between all anchor boxes and the target object in each cluster. The calculation formula is:

[0018] d(box,centroid)=1-Iou(box,centroid)

[0019] Among them, d represents the custom distance, box represents the anchor box, centroid represents the target object, and IOU (box, centroid) represents the intersection-over-union ratio between the anchor box and the target object. The smaller the d value, the greater the probability that the anchor box is the best anchor box in the cluster.

[0020] According to the present invention, a method for identifying point symbols in a scanned topographic map is performed using a deep convolutional neural network. Further, a YOLOv4 network includes CSPDarknet53, PANet and a detection head; an ASPP module is introduced between CSPDarknet53 and PANet, and the module extracts multi-scale feature information through multiple parallel convolutional layers using different expansion rates.

[0021] According to the present invention, the method for identifying point symbols in a scanned topographic map using a deep convolutional neural network further includes:

[0022] Cut the entire topographic map into fixed sizes and save the resulting small images in the image folder;

[0023] Traverse the images in the image folder one by one and enlarge their size by 2 times;

[0024] The ASPP-YOLOv4 model recognizes all the enlarged images one by one and stores the detection frame coordinates after conversion according to the original image;

[0025] The recognized images are stitched together and the detection boxes, symbol types, and confidence levels are drawn.

[0026] According to the present invention, a method for identifying point symbols in a scanned topographic map is performed using a deep convolutional neural network. Furthermore, a frozen training method is used during model training, and the loss functions used are a cross entropy loss function and an intersection-over-union loss function.

[0027] Furthermore, the present invention also provides a system for scanning point symbol recognition in a topographic map using a deep convolutional neural network, comprising:

[0028] The dataset construction module is used to construct a point symbol dataset suitable for supervised learning and divide the dataset into proportions;

[0029] A data enhancement module is used to expand the data set by using data enhancement methods on the original maps in the data set;

[0030] Anchor box generation module, used to generate anchor boxes suitable for multiple types of point symbols based on k-means++ clustering algorithm;

[0031] Model building module, used to design the ASPP-YOLOv4 model for positioning and detecting multiple types of point symbols;

[0032] The image cutting and detection module is used to process the entire topographic map using the image processing operation of "first cutting, then enlarging, then detecting, and finally splicing";

[0033] The model training and evaluation module is used to train and evaluate the model using loss functions and data sets.

[0034] The beneficial effects achieved by adopting the above technical solution are:

[0035] 1. Construction of point symbol dataset: The present invention constructs a point symbol dataset suitable for supervised learning. The dataset includes scanned topographic maps of a fixed scale. By annotating these maps, the richness and diversity of the dataset are ensured, providing sufficient data support for subsequent model training.

[0036] 2. Design of DCNN model: By introducing the ASPP module and YOLOv4 network, a deep convolutional neural network (DCNN) model is designed for the point symbol recognition task of scanning topographic maps. The introduction of the ASPP module expands the receptive field, thereby more effectively extracting the multi-scale information of the symbol object, and can effectively recognize point symbols of small size, simple shape and close integration with the background, significantly improving the point symbol recognition accuracy, efficiency and positioning accuracy.

[0037] 3. Data enhancement: In order to cope with the cross-domain problems brought about by map data sets of different styles and improve the robustness of the model, in addition to the conventional data enhancement methods of horizontal flipping, vertical flipping, random cropping, and grayscale, the present invention also proposes Gaussian blurring, color jittering, and small-angle rotation, which effectively enhances the model's adaptability to different image styles and changes and improves the model's generalization ability.

[0038] 4. Improved positioning accuracy: In order to improve the positioning accuracy of point symbols, the present invention uses the k-means++ clustering algorithm to generate anchor boxes that are more suitable for point symbol data sets. In addition, in order to solve the problem of slow processing speed of large-scale map models, a method of "cutting the map first, then enlarging, then detecting, and finally splicing" is designed. Through this cutting and splicing method, the symbol recognition problem in large-scale topographic maps can be effectively handled, which improves the operability in practical applications.

[0039] 5. Experimental results and performance evaluation: In the experiment, the DCNN model of the present invention achieved an average precision (mAP) of 98.11% and an average intersection over union (mIoU) of 0.876 on the test set. Compared with traditional object detectors and classical algorithms, the method of the present invention shows higher accuracy and efficiency in both point symbol recognition and positioning accuracy.

[0040] In summary, the present invention not only improves the recognition accuracy and positioning accuracy of point symbols, but also enhances the robustness of the model through innovative deep learning technology and data enhancement technology. It has significant technical advantages and application prospects, and promotes technological progress in the field of geographic information extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings of the embodiments of the present invention, wherein the drawings are only used to illustrate some embodiments of the present invention, but not to limit all embodiments of the present invention thereto.

[0042] Figure 1 It is a schematic diagram of the process of a method for scanning point symbol recognition in a topographic map using a deep convolutional neural network according to an embodiment of the present invention;

[0043] Figure 2 is a target symbol schematic diagram of a point symbol data set according to an embodiment of the present invention;

[0044] Figure 3 is a schematic diagram of a data enhancement method according to an embodiment of the present invention, wherein (a) is an original image, (b) is horizontally flipped, (c) is vertically flipped, (d) is randomly cropped, (e) is grayed, and (f) is rotated at a small angle;

[0045] Figure 4Schematic diagram of the relationship between the number of anchor boxes and avgIoU obtained based on the k-means++ clustering algorithm according to an embodiment of the present invention;

[0046] Figure 5 Schematic diagram of the ASPP-YOLOv4 network structure of an embodiment of the present invention;

[0047] Figure 6 is a flowchart of symbol recognition for an entire topographic map according to an embodiment of the present invention;

[0048] Figure 7 are experimental results of different methods of the embodiments of the present invention on the test set;

[0049] Figure 8 is a statistical graph of AP values ​​of ten point symbols recognized by different models in an embodiment of the present invention;

[0050] Fig. 9 It is a schematic diagram of point symbol recognition results of a large-scale map according to an embodiment of the present invention. DETAILED DESCRIPTION

[0051] The following will be combined with the drawings of specific embodiments of the present invention to clearly and completely describe the exemplary scheme of the embodiment of the present invention. Unless otherwise defined, the technical terms or scientific terms used in the present invention should be the common meanings understood by people with ordinary skills in the field.

[0052] This embodiment discloses a method for identifying point symbols in a scanned topographic map using a deep convolutional neural network. The implementation process is as follows: Figure 1 As shown, the following steps are included:

[0053] Step S1: To support the point symbol recognition task, a point symbol dataset suitable for supervised learning is constructed and the dataset is divided proportionally.

[0054] The constructed dataset consists of two parts: 1909 scanned topographic maps at a scale of 1:50,000 and 2505 data-enhanced maps from the scanned topographic maps (including images of the scanned topographic maps that are enlarged and reduced with scale differences). These images include different types of point symbols, and ten typical point symbols are selected, such as Figure 2 As shown (including marked geographical features, buildings, wellheads, etc.), it covers all the symbols of topographic maps.

[0055] The constructed image dataset is divided into training set, validation set and test set with a ratio of 8:1:1. This ensures that the training data is rich enough and the generalization ability of the model can be evaluated on the validation set and test set.

[0056] Each scanned topographic map in the dataset is annotated with point symbols, and the category of the symbol and its coordinates in the image are recorded. This information provides the necessary data support for the supervised learning of the model during training and testing.

[0057] Step S2: In order to enhance the generalization ability of the model, the original map in the data set is enlarged using data augmentation methods to adapt to the robustness and cross-domain problems in symbol recognition. The data augmentation methods are as follows: Figure 3 shown.

[0058] ① Conventional data enhancement method

[0059] During the training process, common image data augmentation methods are used to enhance the generalization ability of the model, such as random cropping: randomly cropping the original image to increase the diversity of the data. Horizontal and vertical flipping: randomly flipping the image to simulate symbols in different directions. Grayscale conversion: converting color images to grayscale images to enhance the model's adaptability to different color styles.

[0060] ②Affine transformation (small angle rotation)

[0061] In order to simulate the small angle tilt that may occur during the scanning process, affine transformation is used for data enhancement. By rotating the scanned topographic map at a small angle, possible distortion in the actual scanned image is simulated, and the model's recognition ability for the rotated image is enhanced.

[0062] ③Gaussian blur and color jitter

[0063] A combined data augmentation method of Gaussian blur and color jittering is applied to scanned topographic maps to simulate map images of different styles and improve the model's adaptability to style changes.

[0064] Step S3: In order to improve the positioning accuracy, an anchor box suitable for multiple types of point symbols is generated based on the k-means++ clustering algorithm.

[0065] Since the sizes of the annotation bounding boxes of point symbols vary greatly, the anchor boxes preset by YOLOv4 may not be applicable to this dataset. Therefore, it is necessary to generate more suitable anchor boxes for point symbols.

[0066] The k-means++ clustering algorithm is used to obtain multiple clusters through iterative calculation according to the anchor box size of the symbols in the data set. The algorithm selects the optimal anchor box size for each cluster by calculating the intersection ratio between all anchor boxes and the target object in each cluster. The calculation formula is:

[0067] d(box,centroid)=1-IoU(box,centroid)

[0068] Among them, d represents the custom distance, box represents the anchor box, centroid represents the target object, and IOU (box, centroid) represents the intersection-over-union ratio between the anchor box and the target object. The smaller the d value, the greater the probability that the anchor box is the best anchor box in the cluster.

[0069] Through experiments, the experimental results are as follows Figure 4 As shown, 9 anchor boxes were finally selected (in Figure 2 The Chinese scientific observatory and the water plant share an anchor frame of the same size) to obtain the best positioning accuracy, with specific values ​​of [(19×40), (33×34), (26×54), (43×42), (37×67), (51×52), (62×61), (53×92), (79×78)].

[0070] Step S4: Design the ASPP-YOLOv4 model to locate and detect multiple types of point symbols.

[0071] The ASPP-YOLOv4 model structure is as follows Figure 5 As shown in the figure, YOLOv4 is used as the basic framework of the target detection model. The YOLOv4 network includes CSPDarknet53, PANet and detection head. YOLOv4 can efficiently detect and locate multiple types of objects. However, since the point symbols of the scanned topographic map are small and closely integrated with the background, the default anchor box and feature extraction method of YOLOv4 are difficult to effectively identify these small objects.

[0072] In order to deal with the problem of low accuracy in small object detection, the Atrous Spatial Pyramid Pooling (ASPP) module is introduced between CSPDarknet53 and PANet. This module can extract multi-scale feature information through multiple parallel convolutional layers with different dilation rates, reducing the problem of detail loss caused by low resolution of feature maps. The ASPP module improves the model's ability to detect small-sized symbols. The Path Aggregation Network (PANet) further improves the model's ability to fuse features of symbols of different scales. Through multi-layer feature fusion, the model can better understand the global and local information in the image.

[0073] The output of the ASPP-YOLOv4 model consists of three parts: ① Bounding box regression: predicting the bounding box position of the point symbol; ② Classification prediction: predicting the category of each point symbol; ③ Confidence: predicting the confidence of the point symbol.

[0074] The DIoU non-maximum suppression (NMS) algorithm is used to further improve the precise positioning and deduplication effects of point symbols.

[0075] Step S5: For large-size topographic maps, the image processing operation of "first cutting, then enlarging, then detecting, and finally splicing" is adopted. The flowchart is as follows: Figure 6 As shown, the specific process is as follows:

[0076] Step S501, cut the entire topographic map into a size of 208×208, fill the pixels of the part less than 208 by padding with zeros, and record the pixel values ​​filled in the width and height directions, and save the obtained small image in the image folder.

[0077] Step S502, traverse the small images in the image folder one by one, and enlarge their size by 2 times to 416×416 to meet the fixed size required by the model input.

[0078] Step S503: The ASPP-YOLOv4 model recognizes all the enlarged images one by one, and stores the detection frame coordinates after conversion according to the original image.

[0079] Step S504: stitch the recognized images together, and draw the detection frame, symbol type and confidence level.

[0080] Step S6: Use the loss function and data set to train and evaluate the model.

[0081] (1) Training process

[0082] a. The public PASCAL VOC2007+2012 dataset is selected for training on the YOLOv4 model. The dataset contains 240,000 labeled images. Large-scale data can enable the model to better learn the shallow features of the images.

[0083] b. Due to the generalization ability of the model, the weight parameters pre-trained on a large-scale dataset can be loaded into the YOLOv4 backbone feature extraction network CSPDarknet module, and other layers use randomly initialized weights.

[0084] c. Use the frozen training method to fix the backbone network, and use the constructed point symbol dataset to fine-tune only some network layers by setting the learning rate. The fine-tuned parameters will be updated with the back-propagation process.

[0085] d. Finally, unfreeze training is performed to continuously adjust all layer parameters through back propagation, and finally obtain the appropriate parameter matrix and bias vector.

[0086] e. The weighted cross entropy loss function and CIoU loss function are used in the training process to ensure the accuracy of symbol recognition and positioning accuracy. The complete intersection over union (CIoU) loss function is adopted to ensure the accuracy of the bounding box is optimized during training.

[0087] (2) Evaluation indicators

[0088] When evaluating the symbol recognition results, important evaluation indicators include positioning evaluation, accuracy evaluation and speed evaluation. This scheme uses the mean of Itersection over Union (mIoU) value of each category as the positioning evaluation indicator. The larger the value, the more accurate the model predicts the symbol location. The calculation formula is as follows:

[0089]

[0090] For a certain symbol category, the true symbol bounding box is called a positive sample (Positive), and the non-symbol bounding box is called a negative sample (Negative). Then the confusion matrix of this symbol can be obtained, as shown in the following table. According to the confusion matrix, the precision and recall can be calculated as follows:

[0091]

[0092]

[0093] Among them, TP, FP, and FN represent true positive, false positive, and false negative, respectively.

[0094] (3) Performance evaluation

[0095] When evaluated on the test set, the experimental results are as follows Figure 7 As shown in the figure, the model based on the present invention performs well in the point symbol recognition and positioning tasks, achieving an average precision (mAP) of 98.11% and an average intersection over union (mIoU) of 0.876. Compared with traditional target detection methods, the proposed DCNN model has obvious advantages in small object detection and symbol positioning. The AP values ​​of ten point symbols recognized by different models are shown in the figure. Figure 8 As shown, it can be seen that the method of the present invention has obvious advantages.

[0096] In addition, the present invention adopts the sample map example of "GBT 20257.3-2017 National Basic Scale Map Format Part 3: 1:25000 1:50000 1:100000 Topographic Map Format" Figure B.8 Village (II) as the base map, and randomly places the ten types of point symbols studied into the image. The scanned image is used as an experimental sample to verify the feasibility of the whole topographic map recognition idea of ​​"first cutting the image, then enlarging, then detecting, and finally splicing" proposed in this invention. The test results are as follows Fig. 9As shown. All target point symbols in this complete topographic map are correctly identified, and the image recognition accuracy of the tilted image during the scanning process is also very high, which is significantly more advanced than the recognition by compressing the size of a single image. Therefore, it is feasible to apply the method proposed in the present invention to actual production operations, which can effectively improve efficiency and ensure accuracy.

[0097] Experimental results show that the deep learning method is significantly better than the traditional algorithm in terms of accuracy and efficiency of point symbol recognition. The introduction of ASPP significantly improves the average precision and intersection over union (IoU) of the model. The data enhancement method effectively alleviates the cross-domain problem and improves the robustness of the model. This paper provides strong support for the research and evaluation of the extraction algorithm of geographic elements in scanned topographic maps.

[0098] Corresponding to the above method, this embodiment also discloses a system for scanning point symbol recognition in a topographic map using a deep convolutional neural network, comprising:

[0099] The dataset construction module is used to construct a point symbol dataset suitable for supervised learning and divide the dataset into proportions.

[0100] The data enhancement module is used to expand the data set by using the data enhancement method on the original map in the data set.

[0101] The anchor box generation module is used to generate anchor boxes suitable for multiple types of point symbols based on the k-means++ clustering algorithm.

[0102] Model building module, used to design the ASPP-YOLOv4 model for positioning and detection of multiple types of point symbols.

[0103] The image cutting and detection module is used to perform image processing operations of "first cutting, then enlarging, then detecting, and finally splicing" on the entire topographic map.

[0104] The model training and evaluation module is used to train and evaluate the model using loss functions and data sets.

[0105] Unless otherwise specifically stated, the components, steps, numerical expressions and values ​​set forth in these embodiments do not limit the scope of the present invention.

[0106] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0107] The units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person of ordinary skill in the art may use different methods to implement the described functions for each specific application, but such implementation is not considered to be beyond the scope of the present invention.

[0108] Those skilled in the art will appreciate that all or part of the steps in the above method can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk or an optical disk. Optionally, all or part of the steps in the above embodiment can also be implemented using one or more integrated circuits, and accordingly, each module / unit in the above embodiment can be implemented in the form of hardware or in the form of software function modules. The present invention is not limited to any specific form of combination of hardware and software.

[0109] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or replace some of the technical features therein by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for identifying point symbols in a scanned topographic map using a deep convolutional neural network, characterized in that: include: Construct a point symbol dataset suitable for supervised learning and divide the dataset proportionally; Use data augmentation methods to expand the dataset for the original maps in the dataset; Generate anchor boxes suitable for multiple types of point symbols based on the k-means++ clustering algorithm; Design the ASPP-YOLOv4 model to locate and detect multiple types of point symbols; The image processing operation of "cutting the map first, then enlarging it, then detecting it, and finally splicing it" is adopted for the whole topographic map; Use loss functions and datasets to train and evaluate the model.

2. The method for identifying point symbols in a scanned topographic map using a deep convolutional neural network according to claim 1, characterized in that: The constructed point symbol dataset includes fixed-scale scanned topographic maps; each scanned topographic map in the dataset is annotated with point symbols, and the symbol category and its coordinates in the image are recorded.

3. The method for identifying point symbols in a scanned topographic map using a deep convolutional neural network according to claim 1, characterized in that: The data enhancement method includes horizontal flipping, vertical flipping, random cropping, graying, Gaussian blurring, color jittering and small-angle rotation of the original image.

4. The method for identifying point symbols in a scanned topographic map using a deep convolutional neural network according to claim 1, characterized in that: The anchor boxes generated based on the k-means++ clustering algorithm for multiple types of point symbols include: The k-means++ clustering algorithm is used to obtain multiple clusters through iterative calculation according to the anchor box size of the symbols in the data set; the algorithm selects the optimal anchor box size for each cluster by calculating the intersection ratio between all anchor boxes and target objects in each cluster. The calculation formula is: d(box, centroid) = 1-IoU(box, centroid), where d represents the custom distance, box represents the anchor box, centroid represents the target object, and IOU(box, centroid) represents the intersection-over-union ratio between the anchor box and the target object. The smaller the d value, the greater the probability that the anchor box is the best anchor box in the cluster.

5. The method for identifying point symbols in a scanned topographic map using a deep convolutional neural network according to claim 1, characterized in that: The YOLOv4 network includes CSPDarknet53, PANet and detection head; the ASPP module is introduced between CSPDarknet53 and PANet, which extracts multi-scale feature information through multiple parallel convolutional layers with different expansion rates.

6. The method for identifying point symbols in a scanned topographic map using a deep convolutional neural network according to claim 1, characterized in that: The image processing operations of "cutting the map first, then enlarging it, then detecting it, and finally splicing it" for the entire topographic map include: Cut the entire topographic map into fixed sizes and save the resulting small images in the image folder; Traverse the images in the image folder one by one and enlarge their size by 2 times; The ASPP-YOLOv4 model recognizes all the enlarged images one by one and stores the detection frame coordinates after conversion according to the original image; The recognized images are stitched together and the detection boxes, symbol types, and confidence levels are drawn.

7. The method for identifying point symbols in a scanned topographic map using a deep convolutional neural network according to claim 1, characterized in that: The frozen training method is used for model training, and the loss functions used are the cross entropy loss function and the intersection-over-union loss function.

8. A system for scanning point symbol recognition in topographic maps using a deep convolutional neural network, characterized in that: include: The dataset construction module is used to construct a point symbol dataset suitable for supervised learning and divide the dataset into proportions; A data enhancement module is used to expand the data set by using data enhancement methods on the original maps in the data set; Anchor box generation module, used to generate anchor boxes suitable for multiple types of point symbols based on k-means++ clustering algorithm; Model building module, used to design the ASPP-YOLOv4 model for positioning and detecting multiple types of point symbols; The image cutting and detection module is used to process the entire topographic map using the image processing operation of "first cutting, then enlarging, then detecting, and finally splicing"; The model training and evaluation module is used to train and evaluate the model using loss functions and data sets.

9. A computer device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images