DOM ground object automatic extraction method and system based on key point detection
By using YOLOv8 key point network for cutting and chunking parallel detection in DOM land object extraction, the problems of low accuracy and low efficiency in the existing technology are solved, and high-precision and fast land object extraction are achieved, suitable for high-precision positioning and Drone aerial survey and other fields.
Patent Information
- Application Number
- CN202411250367.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-06
- Publication Date
- 2024-12-13
AI Technical Summary
The existing technology has problems of low accuracy and low efficiency in DOM land objects extraction. The traditional manual mapping method is costly and low efficiency. The existing artificial intelligence recognition algorithm is not very accurate for high-resolution data extraction in large scenarios and needs to rely on all Manual or semi-automatic semi-manual repair.
The object detection model based on YOLOv8 key point network is adopted to crop and block the DOM graph and perform parallel detection. The key point position is predicted through the convolutional neural network feature extraction and regression algorithm, and high-precision object detection box and key point coordinates are generated. Realize land vector extraction.
It greatly improves the accuracy and efficiency of land objects extraction, reduces manual intervention, and can quickly and automatically extract the refined corner points of land objects and generate high-precision vectorized contours, which are suitable for high-precision positioning and drone aerial surveying and other fields.
Smart Images
Figure CN119131612A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of remote sensing image processing, and more specifically, to a method and system for automatically extracting DOM objects based on key point detection. Background Art
[0002] DOM (Digital Orthophoto Map) is a digital orthophoto dataset generated by using the Digital Elevation Model (DEM) to correct the projection difference of scanned digitized aerial photographs pixel by pixel, and then cutting them according to the scope of the national basic scale topographic map. Remote sensing images contain rich information about objects. How to quickly and accurately identify and extract the information about objects in remote sensing images has become a huge challenge facing the current application of remote sensing images. For example, the refined vector extraction of objects (such as buildings, roads, farmland, photovoltaic panels) in the city-level surveying and mapping results DOM involves the vital interests of residents, so the accuracy of the extracted object vectors and the production efficiency of surveying and mapping results are very high.
[0003] The traditional method of extracting and identifying ground object vectors is mainly the manual mapping method, that is, manually re-mapping the corresponding ground objects from the remote sensing image. Although the accuracy is high, the workload is very large, requiring a lot of manpower and material resources, high cost and low efficiency. At present, there are also proposals to use typical artificial intelligence recognition algorithms such as U-Net, Mask-RCNN, GeoSAM, etc. to extract ground object vectors, but there is a problem of low extraction accuracy for large-scale high-resolution DOM data. In actual production, it can only rely on full manual fine vectorization or semi-automatic and semi-manual repair methods. Summary of the invention
[0004] In order to overcome the defects of low precision and low efficiency of the DOM feature extraction method described in the prior art, the present invention provides a DOM feature automatic extraction method and system based on key point detection.
[0005] In order to solve the above technical problems, the technical solution of the present invention is as follows:
[0006] A method for automatically extracting DOM objects based on key point detection includes the following steps:
[0007] Get the DOM graph to be tested;
[0008] The DOM image is cut and divided into blocks according to a preset block size and overlap to obtain a block image;
[0009] Input the block images into a target detection model based on a YOLOv8 key point network to obtain target detection frames and key point coordinates of each block image;
[0010] The detection results of each block image are vector-merged to generate a ground object extraction vector map with target detection boxes and key point coordinates.
[0011] Furthermore, the present invention also proposes a DOM feature automatic extraction system based on key point detection, which applies the DOM feature automatic extraction method proposed in the present invention. The system comprises:
[0012] Image acquisition module, used to collect the DOM image to be tested;
[0013] An image processing module is used to cut and divide the DOM image into blocks according to a preset block size and overlap to obtain a block image;
[0014] A ground object detection module, on which a pre-trained target detection model based on a YOLOv8 key point network is mounted; the ground object detection module is used to input the block image into the target detection model based on the YOLOv8 key point network to obtain the target detection frame and key point coordinates of each block image;
[0015] The output module is used to vector-merge the detection results of each block image, generate and output the ground object extraction vector map with target detection frame and key point coordinates.
[0016] Furthermore, the present invention also proposes a device, including a memory and a processor, wherein the memory stores computer-readable instructions, wherein when the computer-readable instructions are executed by the processor, the processor executes the steps of the DOM feature automatic extraction method as described in the present invention.
[0017] Furthermore, the present invention also proposes a storage medium on which computer-readable instructions are stored, wherein the computer-readable instructions, when executed by a processor, implement the steps of the method for automatically extracting DOM features as described in the present invention.
[0018] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:
[0019] The present invention adopts the YOLOv8 key point network to build a target detection model. The principle is to use a convolutional neural network to extract features from the image, and then predict the position of the key points and the target detection frame through a regression algorithm to obtain a high-precision target detection frame and key point coordinates. It can quickly and automatically extract the refined corner points of the specified object and generate a high-precision vectorized contour, providing high-quality and high-precision surveying and mapping results, greatly reducing manual intervention;
[0020] The present invention also cuts and divides the DOM into blocks, and inputs the block images into the target detection model for parallel detection, thereby greatly improving the detection speed and realizing rapid identification and extraction of key points of the target object. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1 The figure is a flow chart of a method for automatically extracting DOM features according to an embodiment of the present invention.
[0022] Figure 2 FIG. 4 is a schematic diagram of DOM image segmentation according to an embodiment of the present invention.
[0023] Figure 3 The diagram is an architecture diagram of a YOLOv8 key point network according to an embodiment of the present invention.
[0024] Figure 4 FIG. 1 is a schematic diagram of dilated convolutions with different dilation rates according to an embodiment of the present invention.
[0025] Figure 5 The figure is a diagram showing the architecture of a dilated spatial pyramid pooling layer according to an embodiment of the present invention.
[0026] Figure 6 This is a comparison chart of the accuracy of the target detection model improved based on YOLOv8 according to one embodiment of the present invention.
[0027] Figure 7 The figure is a schematic diagram of merging target detection frames according to an embodiment of the present invention.
[0028] Figure 8 The figure is a schematic diagram of removing abnormal points and surfaces according to an embodiment of the present invention.
[0029] Fig. 9 A vector diagram of target objects and key point extraction is shown according to an embodiment of the present invention.
[0030] Fig.10 The figure is a schematic diagram of a system for automatically extracting DOM features according to an embodiment of the present invention. DETAILED DESCRIPTION
[0031] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.
[0032] The terms used in the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The singular forms "a", "the" and "the" used in the present invention and the appended claims are also intended to include plural forms unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.
[0033] It should be understood that although the terms first, second, third, etc. may be used in the present invention to describe various information, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present invention, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".
[0034] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.
[0035] Example 1
[0036] This embodiment proposes a DOM feature automatic extraction method based on key point detection, such as Figure 1 FIG. 1 is a flow chart of the method for automatically extracting DOM features according to the present embodiment.
[0037] The method for automatically extracting DOM objects based on key point detection proposed in this embodiment includes the following steps:
[0038] S1. Get the DOM graph to be tested;
[0039] S2, cutting and dividing the DOM image into blocks according to preset block sizes and overlaps to obtain block images;
[0040] S3, inputting the block images into a target detection model based on a YOLOv8 key point network to obtain target detection frames and key point coordinates of each block image;
[0041] S4. Perform vector merging on the detection results of each block image to generate a ground object extraction vector map with a target detection frame and key point coordinates.
[0042] In this embodiment, the YOLOv8 key point network is used to build a target detection model. The principle is to use a convolutional neural network to extract features from the image, and then predict the position of the key points and the target detection frame through a regression algorithm to obtain a high-precision target detection frame and key point coordinates. It can quickly and automatically extract the refined corner points of the specified object and generate a high-precision vectorized contour, providing high-quality and high-precision surveying and mapping results, and greatly reducing manual intervention.
[0043] This embodiment also cuts and divides the DOM into blocks, and inputs the block images into the target detection model for parallel detection, thereby greatly improving the detection speed and achieving rapid identification and extraction of key points of the target object.
[0044] This embodiment can be applied to automated production in the fields of high-precision positioning, drone aerial survey, etc., to improve production efficiency and the quality of production results.
[0045] In an optional embodiment, in step S2, the method further comprises the following steps:
[0046] A GPU parallel acceleration strategy is adopted, and the target detection model is used to perform multi-threaded target detection on the block image.
[0047] This embodiment divides the image blocks with overlapping edges of the high-resolution large-scene DOM, adopts multi-threading and GPU parallel acceleration detection, quickly identifies and extracts objects and their refined contour points, and can perfectly support the processing of huge volumes of high-resolution large-scene DOM data, improving the overall object recognition and key point automatic extraction efficiency.
[0048] For example, Figure 2 The figure shows a schematic diagram of DOM image segmentation in this embodiment.
[0049] In an optional embodiment, the target detection model adopts a YOLOv8 key point network, whose backbone network and neck network include several multi-scale dilated convolution modules; the pooling layer in the neck network uses an atrous spatial pyramid pooling layer (Atrous Spatial Pyramid Pooling, ASPP); the output end of the atrous spatial pyramid pooling layer is connected to the head network of the target detection model.
[0050] Among them, the target detection model is built based on the YOLOv8 key point network, which consists of a backbone network (Backbone), a neck network (Neck) and a head network (Head); the backbone network is used to extract features from the input image and convert the image into a feature representation with rich semantic information; the neck network is used to fuse the features output by the backbone network to improve the performance of the model, and to use pooling operations of different scales to improve the detection ability of targets of different sizes; the head network is used as an inference layer to generate detection results, including predicting the bounding box regression value of each anchor box and the confidence of the target existence.
[0051] For example, Figure 3 , which is an architecture diagram of the YOLOv8 key point network of this embodiment.
[0052] The neck network and the head network form a regression prediction network, which is used to perform regression prediction on the features output by the backbone network and predict the target detection frame and key point coordinates of the input image.
[0053] Exemplarily, the fully connected layer in the target detection model includes a C2f module.
[0054] Exemplarily, the dilated spatial pyramid pooling layer is spliced with the skip connection structure in the YOLOv8 network, and the feature map output by the dilated spatial pyramid pooling layer deployed in the neck network is transmitted to the head network for target detection.
[0055] In an optional embodiment, the dilated convolution module includes 6 parallel levels; wherein:
[0056] The first level outputs the input feature map as the first convolution result;
[0057] The second layer performs a 3×3 dilated convolution operation on the input feature map to obtain the second convolution result.
[0058] The third level fuses the input feature map with the second convolution result, and then performs a 3×3 dilated convolution operation with a dilation rate of 6 to obtain the third convolution result.
[0059] The fourth level fuses the input feature map with the second and third convolution results, and then performs a 3×3 dilated convolution operation with a dilation rate of 12 to obtain the fourth convolution result.
[0060] The fifth level fuses the input feature map with the second convolution result, the third convolution result, and the fourth convolution result, and then performs a 3×3 hole convolution operation with a dilation rate of 18 to obtain the fifth convolution result.
[0061] The sixth level fuses the input feature map with the second convolution result, the third convolution result, the fourth convolution result, and the fifth convolution result, and then performs a 3×3 hole convolution operation with a dilation rate of 24 to obtain the sixth convolution result;
[0062] The convolution results of all levels are fused and the first feature map of multi-scale features is output.
[0063] This embodiment adds multi-scale hole convolution on the basis of the YOLOv8 key point network, and adds a pyramid pooling layer to improve the key point extraction network training and reasoning effect of high-resolution large-scene DOM, and improve the accuracy of DOM object contour and key point extraction.
[0064] The principle of the atrous convolution module added in this embodiment is to add a dilation rate to the convolution layer. The dilation rate is the spacing between values when the convolution kernel processes data. Under different dilation rates, the receptive field is also different.
[0065] like Figure 4 The figure shows a schematic diagram of dilated convolutions with different dilation rates. Dilated convolutions can expand the feature receptive field without being affected by the feature spatial resolution. Each convolution can output feature information in a wider range than ordinary convolutions. Dilated convolutions with different dilation rates can be combined to obtain multi-scale features, thereby adapting to the feature structure of high-resolution large-scene images.
[0066] Exemplarily, the expression of the dilated convolution is:
[0067]
[0068] Among them, Y(i,j) represents the pixel value of the i-th row and j-th column of the convolution result, M and N represent the height and width of the convolution kernel, the convolution kernel is K, the expansion factor is K, and X(i,j) represents the pixel value of the i-th row and j-th column of the input data.
[0069] The dilated convolution module in this embodiment uses multi-scale dilated convolutions to combine multiple dilated convolutions with different expansion rates. The later neurons have larger receptive fields, allowing the network to obtain multi-scale features of images in a larger range.
[0070] Furthermore, this embodiment uses a dilated spatial pyramid pooling layer to acquire high-resolution global information and low-resolution local detail information by integrating multi-scale features, thereby improving the accuracy of ground object extraction at different sizes.
[0071] In an optional embodiment, the atrous spatial pyramid pooling layer is spliced with a skip connection structure in a target detection model based on a YOLOv8 key point network; wherein the atrous spatial pyramid pooling layer includes a coding layer, a pyramid pooling layer, and a decoding layer connected in sequence.
[0072] The encoding layer extracts features from the feature map output by the skip connection structure, performs maximum pooling processing after 4 downsampling, and then inputs the feature map into the pyramid pooling layer;
[0073] The pyramid pooling layer averages the input feature map at four scales to obtain feature maps at four scales.
[0074] The decoding layer upsamples the feature maps of the four scales output by the pyramid pooling layer four times, and then restores the image to the original image size layer by layer through transposed convolution, and outputs a second feature map that integrates multi-scale features.
[0075] like Figure 5 , which is an architecture diagram of the atrous spatial pyramid pooling layer of this embodiment.
[0076] Exemplarily, the pyramid pooling layer averages the input feature map at four scales, namely, 1×1, 2×2, 4×4, and 8×8, to obtain four feature maps of corresponding scales.
[0077] For example, the recognition accuracy of the improved target detection model based on YOLOv8 is compared with that of the original YOLOv8 network, as shown in Figure 6 As shown in FIG. 1 , it is a comparison chart of the accuracy of the target detection model improved based on YOLOv8 in this embodiment. Figure 6 Part (a) is a comparison chart of target recognition accuracy. Figure 6 Part (b) is a comparison chart of key point recognition accuracy. Experiments show that the accuracy of target recognition is improved by training with the improved network. The training accuracy of the original network for high-resolution large-scene DOM is low. The accuracy of the improved training network is significantly improved and the fluctuation is more stable.
[0078] In this embodiment, the YOLO v8 key point network is transformed by multi-scale hole convolution to expand the receptive field, so that each convolution output contains feature information in a larger range than ordinary convolution, which is conducive to obtaining global information of ground object features in remote sensing images; at the same time, a pyramid pooling layer is added in combination with the skip connection structure in the YOLO v8 key point network to integrate features of more scales to obtain high-resolution global overall information and low-resolution local detail information, so as to achieve high-precision extraction of ground objects.
[0079] Further, in an optional embodiment, the detection results of each block image are vector-merged in step S4, including the following steps:
[0080] S4.1. According to the target detection frames of each block image, each target detection frame is subjected to intersection detection, and the intersecting target detection frames are merged; the expression is:
[0081]
[0082] Among them, merged_polygons is the set of merged polygon detection boxes; Indicates merging intersecting detection frames, and retaining those that do not intersect with any detection frame directly;
[0083] S4.2, according to the key point coordinates of each block image, the key points within the polygon detection frame are deduplicated, and the key points outside the polygon detection frame are removed;
[0084] S4.3. Perform topological check and error elimination on each object according to the number of key points and connection order; the expression is:
[0085] valid=(N points ==k)∧(I==[0,1,2,...,k-1])
[0086] Among them, valid is the mark of whether it is correct; N points is the number of key points, I is the key point connection order, k is the preset number of key points, and [0,1,2,...,k-1] represents the preset key point order. When the number and order of key points are satisfied at the same time, it is the correct result. Otherwise, it means the key point is wrong or the topology is wrong, and the corresponding key point is removed as an abnormal result.
[0087] S4.4. Connect the remaining key points according to the key point sequence, and vectorize all key points and connecting lines to generate a ground feature extraction vector map with target detection boxes and key point coordinates.
[0088] This embodiment performs vector merging of the target recognition and key point extraction results of each block image, and at the same time eliminates erroneous detection results to output a high-quality vector graph. Since the neighbor blocks are overlapped, the detection frames of the neighbor blocks also overlap. In step S4.1 of this embodiment, each detection frame is intersected and merged, so that the same ground object target elements across multiple blocks are merged to obtain a polygon detection frame set. After the intersection detection and merging, there are no more detection results divided by multiple blocks.
[0089] For example, Figure 7 As shown, it is a schematic diagram of target detection frame merging in this embodiment.
[0090] Furthermore, since the target detection model has specified the key point sequence, multiple repeated key points of the same ground object have the same sequence number. Step S4.2 of this embodiment performs key point deduplication. Specifically, the key points within the detection frame are deduplicated according to the sequence number, and the key points outside the detection frame are directly eliminated as invalid points.
[0091] After merging the detection frames and removing duplicate key points, the detection frames and corresponding key point coordinates of all DOM targets are obtained. Considering possible reasoning errors, step S4.3 of this embodiment performs topological checking and error elimination on each ground object target.
[0092] For example, Figure 8 FIG. 1 is a schematic diagram of abnormal point and surface removal in this embodiment.
[0093] Finally, the key points are connected according to the order of the remaining key points, all key points and connecting lines are vectorized, and the vector graph is output to obtain the final result. Fig. 9 As shown, it is the target object and key point extraction vector diagram of this embodiment.
[0094] This embodiment filters out abnormal points and surfaces based on the number, sequence and topological relationship of key points to improve the accuracy of vector results. In addition, this embodiment uses the refined key points of objects extracted in blocks to quickly merge and generate vector results without any human intervention.
[0095] Furthermore, in an optional embodiment, the method further comprises the following steps:
[0096] Get the training image set;
[0097] The SAM (Segment Anything Model) target recognition model is used to automatically extract and detect each training image to obtain the detection frame and key point position of the target object in each training image;
[0098] Select a portion of training images and input them into the target detection model for training to obtain an initial model;
[0099] The remaining training images are input into the initial model for target detection to obtain detection results of detection frames and key point coordinates, and the correct detection frames and key point positions of the target objects are retained through manual inspection and correction to obtain a complete labeled training image set for training the target detection model.
[0100] In this embodiment, when pre-training the target detection model, a large number of DOM images need to be obtained and annotated to form a training data set. In the traditional key point data annotation method, each target object in each image needs to be manually annotated with target contours and key points, which is labor-intensive and inefficient.
[0101] This embodiment uses the SAM model to assist manual labeling, and a semi-automatic labeling method that trains a preliminary model and then verifies it manually greatly improves the efficiency and quality of data set labeling. It further specifies the number of key points and the order of key points for training, and obtains a target detection model that completes pre-training of various types of ground objects.
[0102] Example 2
[0103] This embodiment proposes a DOM feature automatic extraction system based on key point detection, and applies the DOM feature automatic extraction method proposed in Embodiment 1. Fig.10 , which is a schematic diagram of the automatic DOM feature extraction system of this embodiment.
[0104] The DOM ground feature automatic extraction system based on key point detection proposed in this embodiment includes:
[0105] Image acquisition module, used to collect the DOM image to be tested;
[0106] An image processing module is used to cut and divide the DOM image into blocks according to a preset block size and overlap to obtain a block image;
[0107] A ground object detection module, on which a pre-trained target detection model based on a YOLOv8 key point network is mounted; the ground object detection module is used to input the block image into the target detection model based on the YOLOv8 key point network to obtain the target detection frame and key point coordinates of each block image;
[0108] The output module is used to vector-merge the detection results of each block image, generate and output the ground object extraction vector map with target detection frame and key point coordinates.
[0109] It can be understood that the system of this embodiment corresponds to the method of the above-mentioned embodiment 1, and the options in the above-mentioned embodiment 1 are also applicable to this embodiment, so they will not be described repeatedly here.
[0110] Example 3
[0111] This embodiment proposes a computer device, including a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes all or part of the steps of the DOM feature automatic extraction method proposed in Embodiment 1.
[0112] Example 4
[0113] This embodiment provides a storage medium on which computer-readable instructions are stored. When the computer-readable instructions are executed by a processor, all or part of the steps of the automatic DOM feature extraction method provided in Embodiment 1 are implemented.
[0114] Exemplarily, the storage medium includes, but is not limited to, a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes.
[0115] Exemplarily, the instructions, programs, code sets or instruction sets may be implemented using conventional programming languages.
[0116] Exemplarily, the processor includes but is not limited to a smart phone, a personal computer, a server, a network device, etc., and is used to execute all or part of the steps of the method for automatically extracting DOM features described in Example 1.
[0117] Each embodiment of the present invention is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device embodiment described above is merely exemplary, in which the modules described as separate components may or may not be physically separated, and the functions of each module can be implemented in the same or one or more software and / or hardware when implementing the scheme of the present invention. It is also possible to select some or all of the modules according to actual needs to achieve the purpose of the scheme of this embodiment.
[0118] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.
Claims
1. A method for automatic extraction of DOM features based on key point detection. It is characterized in that The following steps are involved: Get the DOM graph to be tested; The DOM image is cut and divided into blocks according to a preset block size and overlap to obtain a block image; Input the block images into a target detection model based on a YOLOv8 key point network to obtain target detection frames and key point coordinates of each block image; The detection results of each block image are vector-merged to generate a ground object extraction vector map with target detection boxes and key point coordinates.
2. The method for automatically extracting DOM features according to claim 1, It is characterized in that The backbone network and neck network of the target detection model include several multi-scale dilated convolution modules; the pooling layer in the neck network uses a dilated spatial pyramid pooling layer; and the output end of the dilated spatial pyramid pooling layer is connected to the head network of the target detection model.
3. The method for automatically extracting DOM features according to claim 2, It is characterized in that The dilated convolution module includes 6 parallel levels; The first level outputs the input feature map as the first convolution result; The second layer performs a 3×3 dilated convolution operation on the input feature map to obtain the second convolution result. The third level fuses the input feature map with the second convolution result, and then performs a 3×3 dilated convolution operation with a dilation rate of 6 to obtain the third convolution result. The fourth level fuses the input feature map with the second and third convolution results, and then performs a 3×3 dilated convolution operation with a dilation rate of 12 to obtain the fourth convolution result. The fifth level fuses the input feature map with the second convolution result, the third convolution result, and the fourth convolution result, and then performs a 3×3 hole convolution operation with a dilation rate of 18 to obtain the fifth convolution result. The sixth level fuses the input feature map with the second convolution result, the third convolution result, the fourth convolution result, and the fifth convolution result, and then performs a 3×3 hole convolution operation with a dilation rate of 24 to obtain the sixth convolution result; The convolution results of all levels are fused and the first feature map of multi-scale features is output.
4. The method for automatically extracting DOM features according to claim 2, It is characterized in that The atrous space pyramid pooling layer is spliced with a skip connection structure in a target detection model based on a YOLOv8 key point network; wherein the atrous space pyramid pooling layer includes a coding layer, a pyramid pooling layer, and a decoding layer connected in sequence; The encoding layer extracts features from the feature map output by the skip connection structure, performs maximum pooling processing after 4 downsampling, and then inputs the feature map into the pyramid pooling layer; The pyramid pooling layer performs average pooling of the input feature map at four scales to obtain feature maps at four scales; The decoding layer upsamples the feature maps of the four scales output by the pyramid pooling layer four times, and then restores the image to the original image size layer by layer through transposed convolution, and outputs a second feature map that integrates multi-scale features.
5. The method for automatically extracting DOM features according to claim 1, It is characterized in that The vector merging of the detection results of each block image includes the following steps: According to the target detection frames of each block image, each target detection frame is intersected and detected, and the intersecting target detection frames are merged; the expression is: Among them, merged_polygons is the set of merged polygon detection boxes; Indicates merging intersecting detection frames, and retaining those that do not intersect with any detection frame directly; According to the key point coordinates of each block image, the key points within the polygon detection frame are deduplicated, and the key points outside the polygon detection frame are removed; According to the number of key points and the connection order, each object is topologically checked and errors are eliminated; the expression is: valid=(N points ==k)∧(I==[0,1,2,...,k-1]) Among them, valid is the mark of whether it is correct; N points is the number of key points, I is the key point connection order, k is the preset number of key points, and [0,1,2,...,k-1] represents the preset key point order. When the number and order of key points are satisfied at the same time, it is the correct result. Otherwise, it means the key point is wrong or the topology is wrong, and the corresponding key point is removed as an abnormal result. The remaining key points are connected according to the key point sequence, and all key points and connecting lines are vectorized to generate a ground object extraction vector map with target detection boxes and key point coordinates.
6. The method for automatically extracting DOM features according to any one of claims 1 to 5, It is characterized in that The method further comprises the following steps: Get the training image set; The SAM target recognition model is used to automatically extract and detect each training image, and the detection frame and key point position of the target object in each training image are obtained; Select a portion of training images and input them into the target detection model for training to obtain an initial model; The remaining training images are input into the initial model for target detection to obtain detection results of detection frames and key point coordinates, and the correct detection frames and key point positions of the target objects are retained through manual inspection and correction to obtain a complete labeled training image set for training the target detection model.
7. The method for automatically extracting DOM features according to any one of claims 1 to 5, It is characterized in that The method further comprises the following steps: A GPU parallel acceleration strategy is adopted, and the target detection model is used to perform multi-threaded target detection on the block image.
8. A DOM feature automatic extraction system based on key point detection, applying the DOM feature automatic extraction method according to any one of claims 1 to 7, It is characterized in that include: Image acquisition module, used to collect the DOM image to be tested; An image processing module is used to cut and divide the DOM image into blocks according to a preset block size and overlap to obtain a block image; A ground object detection module, on which a pre-trained target detection model based on a YOLOv8 key point network is mounted; the ground object detection module is used to input the block image into the target detection model based on the YOLOv8 key point network to obtain the target detection frame and key point coordinates of each block image; The output module is used to vector-merge the detection results of each block image, generate and output the ground object extraction vector map with target detection frame and key point coordinates.
9. A device comprising a memory and a processor, wherein the memory stores computer-readable instructions, It is characterized in that When the computer-readable instructions are executed by the processor, the processor executes the steps of the method for automatically extracting DOM features according to any one of claims 1 to 7.
10. A storage medium having computer-readable instructions stored thereon, It is characterized in that When the computer-readable instructions are executed by a processor, the steps of the method for automatically extracting DOM features according to any one of claims 1 to 7 are implemented.