Ground penetrating radar image recognition method based on GDS-YOLOv8n

Abstract

The invention discloses a ground penetrating radar image recognition method, and particularly relates to a ground penetrating radar image recognition method based on GDS-YOLOv8n, and the method comprises the specific steps: 1, constructing a ground penetrating radar image data set, and dividing the data set into a training set, a verification set and a test set; 2, a GDS-YOLOv8n model is constructed, and the GDS-YOLOv8n model is constructed by adopting a DRRB module, an SPD-Conv module, an Inner-SIOU and a detection head network; and step 3, using a GDS-YOLOv8n model to carry out identification on the constructed ground penetrating radar image data set. A DRRB module, an SPD-Conv module, an Inner-SIoU and an auxiliary training head are introduced into a YOLOv8n network model, and an identification model GDS-YOLOv8n of a GPR image is provided. And meanwhile, the GDS-YOLOv8n model subjected to TensorRT optimization is deployed on NVIDIA Jetson Orin NX edge computing equipment, so that the GPR image real-time interpretation of the mobile terminal is realized. The method is excellent in model reasoning speed and resource consumption, can accurately recognize a target in a complex environment, and meets the requirements for real-time performance and portability in practical application.

Description

Technical Field

[0001] The present invention relates to the field of new generation information technology, and in particular to a ground penetrating radar image recognition method based on GDS-YOLOv8n. Background Art

[0002] With the accelerated advancement of urbanization, the development and utilization of urban underground space has been expanding, bringing many conveniences to people's lives. However, with the increase in pipeline laying, subway construction, underground resource development and other projects, coupled with the aging of some underground facilities, urban road collapse, pipeline explosion and other safety accidents have occurred frequently. This will seriously threaten people's lives and property safety. Therefore, it is crucial to accurately and efficiently detect urban underground space targets and understand their types and distribution to ensure the safety of urban underground space.

[0003] Underground target detection technology mainly includes destructive detection and non-destructive detection. Among them, the main destructive detection methods include direct excavation method, earth pit method, trench profile method and soil core method. Although these methods can directly obtain detailed information of underground targets, they have high economic and time costs. In addition, they are also destructive to a certain extent and it is difficult to meet the needs of large-scale rapid detection. In contrast, non-destructive detection methods such as electromagnetic detection, acoustic detection and nuclear radiation detection have been widely used due to their advantages such as non-destructiveness, high efficiency and repeatable detection. Among them, Ground Penetrating Radar (GPR), as one of the electromagnetic detection technologies, has the advantages of high detection accuracy, strong penetration and mature technology compared with other non-destructive detection methods, and is easy to combine with data processing technologies such as artificial intelligence. It has shown excellent application value in the fields of underground target exploration, road and bridge quality inspection and archaeological research.

[0004] As a technology of artificial intelligence, machine learning has powerful data processing and analysis capabilities, as well as advantages such as autonomous learning. Therefore, scholars have carried out research on GPR image detection methods based on machine learning and have achieved fruitful results. Traditional GPR image machine learning detection methods mainly include generalized Hough transform (GHT), support vector machine (SVM), clustering algorithm and BP (Back Propagation) neural network. These methods require artificially designed features, are sensitive to training data, and have limited generalization capabilities, making it difficult to adapt to complex and changeable actual detection scenarios.

[0005] With the continuous development of deep learning technology, it has shown obvious advantages in the fields of image segmentation and target detection. For this reason, scholars have also applied it to GPR image data processing and analysis. This method does not require manual design of features, and can automatically learn and extract complex and diverse features, which is particularly suitable for GPR image detection in complex environments. At present, the mainstream deep learning algorithms used for GPR image detection include Faster R-CNN (Fast Region-based Convolutional Network), SSD (Single Shot MultiBox Detector) and YOLO (You Only Look Once). Among them, Faster R-CNN is a two-stage detection algorithm, and the detection process is relatively time-consuming. YOLO and SSD use end-to-end detection methods, have faster detection speeds, and can meet the needs of real-time detection tasks. In addition, YOLO's recognition accuracy is better than SSD, and it is more suitable for GPR image detection task scenarios. Therefore, scholars have used the YOLO algorithm to carry out a lot of research on GPR image detection and recognition.

[0006] The YOLOv8 model is an improvement on the YOLOv5 model, with higher detection accuracy. Among them, the YOLOv8n model, as the network model with the smallest number of parameters and computational complexity in this series, not only maintains high detection accuracy, but also has a faster detection speed. It is particularly suitable for real-time detection of GPR in mobile scenarios. However, this model also has some shortcomings, such as unsatisfactory detection accuracy for GPR images, especially the risk of missed detection and false detection of small targets. Summary of the invention

[0007] In order to solve the above-mentioned problems in the prior art, the present invention provides a ground penetrating radar image recognition method based on GDS-YOLOv8n to solve the current technical problems.

[0008] The technical solution adopted by the present invention to solve its technical problem is: The present invention provides a ground penetrating radar image recognition method based on GDS-YOLOv8n. The method improves YOLOv8n by using technical means such as DRRB (Dilated Residual Reparam Block) feature extraction module, SPD-Conv (space-to-depth-Conv) downsampling module, auxiliary training head and Inner-SIoU loss function, and proposes a GDS-YOLOv8n model suitable for GPR image detection, aiming to effectively solve the problems of low accuracy and poor small target detection performance of YOLOv8n in GPR data detection.

[0009] To achieve the above purpose, the technical solution adopted by the present invention is as follows: A ground penetrating radar image recognition method based on GDS-YOLOv8n, the specific steps include: Step 1: Construct a GPR image dataset: First, obtain the measured GPR images, process them with Reflexw software and save them in JPEG format. Then use LabelImg software to annotate the images, and finally divide the dataset into training set, validation set and test set.

[0010] Step 2: Construction of GDS-YOLOv8n model: DRRB module, SPD-Conv module and Inner-SIoU are used to replace part of the feature extraction module C2f, part of the downsampling convolution module Conv and the loss function CIoU respectively. At the same time, the auxiliary training head (Aux Head) is introduced into the detection head network (Head) to construct the GDS-YOLOv8n network model; Step 3: Use the GDS-YOLOv8n model to perform recognition processing on the constructed GPR image dataset. In the target recognition process, the input GPR image is first preprocessed to automatically adjust the image size while maintaining the original aspect ratio to ensure the integrity of key information. Subsequently, the model uses a multi-layer convolutional neural network to extract the multidimensional features of the target, including shape, size, position, and contrast. In the detection stage, the model generates a bounding box and a corresponding confidence score for each predicted target. Finally, through non-maximum suppression (NMS) processing, the category, location, and confidence information of each recognized target are output.

[0011] Step 4: Deploy the GDS-YOLOv8n model to the NVIDIA Jetson edge device and use the Jetson device's TensorRT to optimize the model file to further improve the detection speed of the model on the Jetson device. The optimization techniques used include computational graph optimization, layer fusion, and quantization.

[0012] The constructing of the GPR image data set specifically includes: Step (1): Select an actual scene, use GPR to conduct field detection, and obtain several measured GPR images; Step (2): After processing with Reflexw software, save it in .jpg format. Then, use LabelImg tool to annotate the GPR image. Finally, divide the dataset into training set, validation set and test set in the ratio of 7:1:2.

[0013] The GDS-YOLOv8n model consists of three parts. The first part is the backbone network (Backbone), which is used to extract rich features from the input image. It extracts low-level to high-level visual features layer by layer through modules such as DRRB and SPD-Conv. The second part is the neck network (Neck), which is used to fuse and enhance multi-scale features from the Backbone network. It usually contains top-down and bottom-up paths to achieve the interaction of features at different levels. The third part is the detection head network (Head), which is used to perform the final target detection task. It receives fused features of three different resolutions from the Neck network and converts them into specific detection results. The detection head usually includes two main tasks: target classification and bounding box regression. The classification task determines which category the detected object belongs to, while the bounding box regression is used to locate the position and size of the target.

[0014] The specific steps of GPR image data recognition and processing of the GDS-YOLOv8n model are as follows: ① GPR image preprocessing After the picture is input into the model, the image size is automatically adjusted to 640×640 while maintaining the original aspect ratio to ensure the integrity of key information so that the image data can be correctly input into the subsequent deep learning model.

[0015] ② Target feature extraction in GPR images The preprocessed image is input into the Backbone network in the GDS-YOLOv8n model for feature extraction. The C2f and DRRB modules in the Backbone network are responsible for extracting multi-scale features from the input image, and the Conv and SPD-Conv modules perform downsampling to generate feature maps of different resolutions.

[0016] ③ Strengthen feature extraction and feature fusion In the Neck network, FPN and PANet are used to fuse feature maps of different resolutions to make full use of multi-scale features and enhance the model's ability to recognize targets.

[0017] ④Target recognition Three YoloHead detection heads are used to predict the position and category of objects on feature maps with resolutions of 20×20, 40×40, and 80×80.

[0018] The benefits of the present invention are as follows: the DRRB module, SPD-Conv module, Inner-SIoU and auxiliary training head are introduced into the YOLOv8n network model, and a GPR image recognition model GDS-YOLOv8n is proposed. At the same time, the GDS-YOLOv8n model optimized by TensorRT is deployed on the NVIDIA Jetson Orin NX edge computing device to achieve real-time interpretation of GPR images on the mobile terminal. The present invention not only excels in model reasoning speed and resource consumption, but also can accurately identify targets in complex environments, meeting the requirements for real-time and portability in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The above contents and advantages of the present invention will become obvious and easily understood from the description of the embodiments in conjunction with the following drawings, in which: Figure 1 A schematic diagram of a process of a ground penetrating radar image recognition method based on GDS-YOLOv8n according to an embodiment of the present invention; Figure 2 A GDS-YOLOv8n model of a ground penetrating radar image recognition method based on GDS-YOLOv8n according to an embodiment of the present invention; Figure 3 The recognition effect of YOLOv8n on the self-built data set before the improvement of the ground penetrating radar image recognition method based on GDS-YOLOv8n in an embodiment of the present invention; Figure 4 The recognition effect of GDS-YOLOv8n on the self-built data set of the ground penetrating radar image recognition method based on GDS-YOLOv8n in an embodiment of the present invention; Figure 5 A model comprehensive performance comparison of the ground penetrating radar image recognition method based on GDS-YOLOv8n in an embodiment of the present invention; Figure 6 This is the model deployment process of the ground penetrating radar image recognition method based on GDS-YOLOv8n in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] Example 1

[0022] In order to improve the intelligent interpretation effect of GPR images. The present invention proposes a ground penetrating radar image recognition method based on GDS-YOLOv8n. The specific improvements include: first, using the DRRB feature extraction module to replace some C2f modules in the YOLOv8n model, aiming to enhance the model's ability to extract multi-scale features. Secondly, using the SPD-Conv downsampling module to replace the Conv module corresponding to the feature map of 320×320 and below pixels, effectively overcomes the problem of information loss in the downsampling process of GPR images with limited resolution and small targets. At the same time, an auxiliary training module is introduced to improve the detection performance of GPR images without increasing the complexity of the model in the detection stage. Finally, the Inner-SIoU loss function is introduced. On the basis of adding new constraints, an auxiliary bounding box suitable for the current GPR image is generated by a scale factor to improve the accuracy of the prediction box. In addition, in order to meet the needs of real-time detection tasks of GPR underground targets in mobile scenarios, the GDS-YOLOv8n model is deployed on the NVIDIA Jetson Orin NX edge computing device and optimized using TensorRT. After FP16 quantization, the detection speed of the model reaches 40.6FPS. The flowchart of the proposed method is as follows: Figure 1 The specific steps are as follows: 1) Constructing GPR image dataset We selected an actual scene and used the Swedish MALA third-generation ProEX ground penetrating radar to conduct field detection of the scene, obtaining 1,714 measured GPR images, which were annotated using the LabelImg tool. Finally, the dataset was divided into training set, validation set, and test set in a ratio of 7:1:2.

[0023] 2) Construction of GDS-YOLOv8n model The DRRB module, SPD-Conv module and Inner-SIoU are used to replace part of the feature extraction module C2f, part of the downsampling convolution module Conv and the loss function CIoU respectively. At the same time, the auxiliary training head (Aux Head) is introduced into the detection head network (Head). Combining the above process, the GPR image recognition network GDS-YOLOv8n is constructed.

[0024] 3) GPR image recognition based on GDS-YOLOv8n Step (1): After the GPR image is input into the model, the image size is adjusted first, that is, the image resolution is automatically adjusted to 640×640×3. After that, it is automatically input into the network and then downsampled through the Backbone network to obtain features of different scales.

[0025] Step (2): Use the Neck network in the GDS-YOLOv8n model to upsample the features obtained in step (1). Then, fuse the feature maps with resolutions of 20×20, 40×40, and 80×80 in the Backbone network and the Neck network to obtain enhanced features of three different dimensions.

[0026] Step (3): Based on step (2), three YoloHead detection heads are used to predict the position and category of objects on feature maps with resolutions of 20×20, 40×40, and 80×80.

[0027] 4) GDS-YOLOv8n model deployment: The GDS-YOLOv8n model is deployed on the Jetson edge computing device, and TensorRT is used to optimize the model to improve its detection speed.

[0028] Example 2 1) Ground Penetrating Radar Image Recognition Method Based on GDS-YOLOv8n The YOLOv8 model is an improvement on the YOLOv5 model, with higher detection accuracy. Among them, the YOLOv8n model, as the network model with the smallest number of parameters and computation in the series, not only maintains a high detection accuracy, but also has a faster detection speed, which is particularly suitable for real-time detection of GPR in mobile application scenarios. However, this model also has some shortcomings, such as unsatisfactory detection accuracy for GPR images, especially the risk of missed detection and false detection for small targets. GDS-YOLOv8n uses DRRB, SPD-Conv, Inner-SIoU and auxiliary training head modules, which can effectively overcome the above problems. Its structure is as follows: Figure 2 As shown in A. Figure 2 As shown in Figure A, the GDS-YOLOv8n network still consists of three parts: Backbone, Neck, and Head. Compared with the YOLOv8n network, the improvements include: ① Using DRRB modules and SPD-Conv modules to replace part of the feature extraction module C2f and part of the downsampling convolution module Conv, respectively, to improve the feature extraction capabilities of the Backbone and Neck networks and reduce information loss during downsampling; ② Introducing an auxiliary training head in the Head to improve the training effect of the model; ③ Using the Inner-SIoU loss function to replace CIoU, the purpose is to improve the accuracy of the prediction box. Among them, the dotted line represents the path for the auxiliary training head to transmit loss information to the network, and (H, V, C1, C2) represent the length, width, and number of input and output channels of the module output image, respectively. Specific improvement measures will be given in detail below.

[0029] ①Construction of DRRB module In the YOLOv8n model, the C2f module uses 3×3 convolution to extract features. This small-sized convolution kernel has a limited receptive field, resulting in a relatively simple semantic feature of the GPR image extracted by the deep layer of the network. In contrast, the DWR (Dilation-wise Residual) module is composed of dilated convolutions with different dilation rates, which can effectively increase the receptive field and improve the efficiency of extracting multi-scale semantic features. To this end, the present invention uses DWR to replace the feature extraction module C2f. However, the original DWR module will produce a certain grid effect due to the use of discrete sampling, which affects the continuity of the features. To solve this problem, DRB (Dilated Reparam Block) modules of different sizes are introduced into the DWR module to replace the dilated convolutions with dilation rates of 3 and 5 in the DWR module, thereby forming a DRRB module, aiming to improve the ability to extract multi-scale features. The constructed DRRB module is shown as follows Figure 2 As shown in D.

[0030] Depend on Figure 2 As can be seen in Figure D, preliminary feature extraction is first performed through a 3×3 standard convolution, and the feature maps are divided into three groups. Then, different processing methods are used for the three groups of feature maps to improve the ability to extract multi-scale semantic information. Among them, one group uses a standard 3×3 convolution, and the other two groups use DRB modules of different sizes. It can be seen that the DRB module is the key to building a DRRB module. Usually, a DRB module consists of a large-size convolution and multiple dilated convolutions with different dilation rates in parallel. The structure of the 9×9DRB module is as follows Figure 2 As shown in B. Figure 2 As shown in Figure B, it includes a 9×9 standard convolution kernel and four dilated convolutions with different dilation rates. This module uses structural reparameterization to improve the inference speed during inference. The structural reparameterization process is as follows: Figure 2 As shown in C in the figure, during inference, it converts five parallel convolutions into one convolution. This not only reduces the number of model parameters, but also increases the model's running speed. In addition, the dilation rate, convolution kernel size, and number of dilated convolutions in the DRB module are adjustable to meet the needs of different detection tasks.

[0031] When replacing the C2f module with the DRRB module, the parameters of the DRB module are the key to affecting the model detection performance. In order to obtain a DRRB model with excellent performance, the present invention analyzes the impact of the DRB model on the detection performance of the DRRB model on the GPR data set, and determines the parameters of the DRB module based on the detection effect of DRRB. When selecting the parameters of the DRB module, the present invention uses a variety of DRB modules to form different DRRB modules, and the results are shown in Table 2. At this time, the parameters of the DRB module used are shown in Table 1, where size represents the size of the DRB module, and k and d are the convolution kernel size and dilation coefficient of the dilated convolution, respectively. It can be seen from Table 2 that in the feature maps with spatial resolutions of 20×20 and 40×40, the detection effect of the DRRB module is optimal when DRB modules of sizes (7, 9) and (9, 11) are used, respectively.

[0032] Table 1 DRB module parameters size k d 5 3, 3 1, 2 7 5, 5, 3 1, 2, 3 9 5, 5, 3, 3 1, 2, 3, 4 11 5, 5, 3, 3, 3 1, 2, 3, 4, 5 13 5, 7, 3, 3, 3 1, 2, 3, 4, 5 .

[0033] Table 2 DRRB module performance comparison 20×20 40×40 Params / Million mAP50 / % （5,7） （7,9） 2.69 93.9 （5,7） （9,11） 2.699 94 （7,9） （9,11） 2.706 95.3 （7,9） （11,13） 2.717 94.8 （9,11） （11,13） 2.727 94.1 , Note: (x, y) means using DRB modules of size x and y.

[0034] ② Replacement of downsampling module The YOLOv8n model uses a Conv module with a stride of 2 for downsampling. However, for GPR images with limited resolution and small targets, due to limited redundant information, spatial fine-grained information is lost and feature learning is insufficient, which in turn affects the detection accuracy. To this end, the present invention uses the SPD-Conv downsampling module to replace the Conv downsampling module in the feature map with a resolution of 320×320 and below in YOLOv8n. The SPD-Conv module consists of two parts: a space-to-depth (SPD) layer and a non-strided convolution layer. Figure 2 As shown in Figure E, the SPD layer slices the spatial features and transfers them to the channel dimension, preserving the information on the original feature map to the greatest extent and avoiding the information loss problem during downsampling. The non-stride convolution layer extracts features from the output of the SPD layer and effectively reduces the amount of computation by reducing the number of channels.

[0035] ③Introduction of auxiliary training head During the training process of the YOLO8n model, the update of the network weights propagates from the deep layer to the shallow layer. However, as the network depth increases, the gradient signal will gradually attenuate during the propagation process, making it difficult for the shallow network to learn effectively, thus affecting the performance of the overall model. To solve this problem, the present invention introduces an auxiliary training head into the detection head of the YOLOv8n model. The model after adding the auxiliary training head is shown in FIG. Figure 2 As shown in F. During training, in addition to using the main detection head (Lead Head) to transmit loss information, the auxiliary training head will also transmit additional loss information to the middle layer of the network. This training method can effectively alleviate the gradient vanishing problem, improve the learning efficiency of shallow network parameters, and enhance the overall performance of the model in GPR image detection.

[0036] ④ Improvement of loss function YOLOv8n uses CIoU as the bounding box loss function by default. Although this loss function takes into account factors such as the overlapping area between the predicted box and the true box, the center point distance, and the aspect ratio, it ignores the angle difference between the two, resulting in a decrease in the convergence speed and accuracy of the predicted box. To this end, the present invention uses the Inner-SIoU bounding box loss function to replace CIoU, aiming to make up for the shortcomings of CIoU by adding the angle loss between the predicted box and the true box. The specific calculation process is as follows: The first step is to use r and the coordinates of the center point of the real frame ,width and height Calculate the coordinates of the auxiliary real box. That is, calculate the coordinates of the upper left vertex of the auxiliary real box and the coordinates of the lower right vertex The calculation formula of its vertex coordinates is: (1), (2).

[0037] The second step is to use r and the coordinates of the center point of the prediction box ,width ,high , calculate the coordinates of the auxiliary prediction box. That is, calculate the coordinates of the upper left vertex of the auxiliary prediction box and the coordinates of the lower right vertex The calculation process of its vertex coordinates is the same as formulas (1) and (2).

[0038] The third step is to calculate the intersection area (inter) and union area (union) of the auxiliary real box and the auxiliary prediction box, and calculate the intersection and union ratio based on this. The calculation process can be described as: (3), (4), (5).

[0039] The fourth step is to calculate the intersection-over-union ratio based on the area of ​​the predicted box and the area of ​​the real box. , and and Introduction The loss function .final The loss function ( ) is calculated as: (6).

[0040] The specific steps of using the GDS-YOLOv8n model to identify underground targets in GPR images are: ① Target feature extraction in GPR images The feature extraction of the GPR image with an input size of 640×640×3 is performed in the Backbone network, and the feature maps with resolutions of 320×320, 160×160, 80×80, 40×40 and 20×20 are obtained in turn through 5 downsampling. Among them, the feature maps with resolutions of 40×40 and 20×20 are obtained by using the DRRB feature extraction module and the SPD-Conv downsampling module; the feature maps with resolutions of 160×160 and 80×80 are obtained by using the C2f feature extraction module and the SPD-Conv downsampling module.

[0041] ②Multiple feature extraction and feature fusion The initial features extracted by the Backbone network are extracted and fused multiple times. The FPN (Feature Pyramid Network) and PANet (Path Aggregation Network) structures are mainly used to achieve the fusion and enhancement of multi-scale features. FPN gradually enlarges the size of the feature map through upsampling, so that the low-level features also contain rich semantic information; while PANet promotes the efficient transmission of the bottom-level information to the top layer of the network through a bottom-up path.

[0042] ③Target recognition Three YoloHead detection heads are used to process the feature maps with resolutions of 20×20, 40×40 and 80×80 in the Neck network respectively. These detection heads work in parallel and analyze feature maps of different scales at the same time to achieve accurate prediction of target location and category.

[0043] In order to verify the recognition performance of the GDS-YOLOv8n network proposed in the present invention, the recognition performance was verified on the dataset constructed by the present invention and compared with that of the YOLOv8n network. The results are as follows: Figure 3 and Figure 4shown.

[0044] Figure 3 and Figure 4 They are the recognition results of YOLOv8n and GDS-YOLOv8n on the self-built data set of the present invention. Figure 3 and Figure 4 It can be seen that GDS-YOLOv8n can locate and identify common underground targets in the self-built GPR image dataset, and its recognition effect is significantly better than the YOLOv8n network.

[0045] To further verify the detection advantage of the GDS-YOLOv8n model, we compared it with the results of YOLOv5n, YOLOv7-tiny, and YOLOv8n models. The results are shown in Table 3.

[0046] Table 3 Comparative experimental results Model Params / Million GFLOPs P / % R / % mAP50 / % mAP50:90 / % FPS(bs=1) <![CDATA[SSD * ]]> 24.28 176.23 84.03 81.86 82.42 48.3 35.2 YOLOv5n 1.77 4.2 89.3 86 89.7 61.1 129.9 YOLOv5s 7.03 15.8 90.7 90.4 92.3 62.8 123.5 YOLOv7-tiny 6.03 13.1 91.9 92.6 91.2 57.6 80 YOLOv8n 3.01 8.1 93.1 90.1 92.8 60.7 109.9 YOLOv8s 11.13 28.4 94.7 90.7 93.7 62 103.1 YOLOv8m 25.8 78.7 94.1 96.3 96.7 63.3 55.6 GDS-YOLOv8n 4.43 11.2 97.1 96.2 96.9 64.3 97.2 .

[0047] As can be seen from Table 3, the P and mAP50 of the GDS-YOLOv8n model are 97.1% and 96.9% respectively, both higher than other detection models. Although R is slightly lower than YOLOv8m by 0.1%, it is still over 96%. Moreover, Params and GFLOPs are lower than those of models such as YOLOv5s, YOLOv7-tiny, and YOLOv8s. This shows that the model has achieved a reasonable compromise between detection accuracy and model complexity. In addition, when the batch size is 1, the inference speed reaches 97.2FPS, which meets the PC's real-time detection requirements for GPR images.

[0048] In addition, in order to further verify the advantages of its comprehensive detection accuracy, this paper comprehensively analyzes three key indicators, mAP50, FPS and Model Size, from the perspective of practical application. These indicators reflect the detection accuracy, inference speed and resource usage of the model, which are particularly important for evaluating the application effect of the model in actual scenarios. The results are as follows Figure 5 As shown. Figure 5 It can be seen that the GDS-YOLOv8n model has a higher FPS and the highest mAP50, and the model size is also smaller. This shows that the model achieves excellent GPR image target recognition capabilities while maintaining a high detection speed, and also has certain advantages in mobile terminal deployment. Therefore, the GDS-YOLOv8n model proposed in the present invention has better performance in GPR image detection.

[0049] Example 3

[0050] In order to meet the usage requirements in mobile application scenarios, the present invention deploys the GDS-YOLOv8n model on the small and low-power NVIDIA Jetson Orin NX edge device. The deployment process is as follows Figure 6 As shown in the figure, first, the trained .pt model file is converted into an .onnx (open neural network exchange) file on the server. This format can store models obtained under different training frameworks in a unified format, simplifying the model migration process and improving the flexibility of deployment. Secondly, in order to further improve the model detection speed, the .onnx model file is optimized using TensorRT of the Jetson Orin NX device, and finally the .engine file is generated. Among them, the optimization techniques used include computational graph optimization, layer fusion, and quantization. The optimization results can be seen from Table 4. Under the premise of maintaining mAP50 without loss, the model inference speed is increased from 22FPS to 40.6FPS, which exceeds the real-time detection standard of 25FPS in the video surveillance field and can meet the real-time detection requirements of mobile GPR data.

[0051] Table 4 GDS-YOLOv8n model deployment format performance comparison Model Model Size / MB mAP50 / % FPS (bs=1) GDS-YOLOv8n.pt 10.3 96.9 22 GDS-YOLOv8n.onnx 17.1 96.9 6.5 GDS-YOLOv8n.engine (FP32) 19.1 96.9 31.4 GDS-YOLOv8n.engine (FP16) 10.7 96.9 40.6 GDS-YOLOv8n.engine (INT8) 5.6 84.1 46 Note: FP32, FP16, and INT8 represent 32-bit, 16-bit floating point, and 8-bit integer quantization models, respectively.

[0052] Example 4 This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, a ground penetrating radar image recognition method based on GDS-YOLOv8n provided in Embodiment 1 of the present invention is implemented.

[0053] Example 5

[0054] This embodiment provides a computer-readable storage medium having a computer program stored thereon, characterized in that when the program is executed by a processor, a ground penetrating radar image recognition method based on GDS-YOLOv8n provided in Embodiment 1 of the present invention is implemented.

[0055] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0056] The solutions in the embodiments of the present application can be implemented using various computer languages, for example, the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0057] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0058] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0059] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0060] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A ground penetrating radar image recognition method based on GDS-YOLOv8n, characterized in that the steps include: Step 1: Build a ground penetrating radar image dataset, which is divided into a training set, a validation set, and a test set. Step 2: GDS-YOLOv8n model construction: DRRB module, SPD-Conv module, Inner-SIoU and detection head network are used to build the GDS-YOLOv8n model; Step 3: Use the GDS-YOLOv8n model to identify the constructed ground penetrating radar image dataset.

2. The ground penetrating radar image recognition method based on GDS-YOLOv8n according to claim 1, characterized in that: The constructing of the ground penetrating radar image dataset comprises: Step 2.1: Select an actual scene, use ground penetrating radar to conduct field detection, and obtain several measured ground penetrating radar images; Step 2.2: Convert the measured ground penetrating radar images into different formats and divide them into training set, validation set and test set in a ratio of 7:1:

2.

3. The ground penetrating radar image recognition method based on GDS-YOLOv8n according to claim 1, characterized in that: The GDS-YOLOv8n model consists of three parts: The first part is the backbone network, including DRRB and SPD-Conv modules, which is used to extract low-level to high-level visual features from the input image layer by layer; The second part is the neck network, which is used to fuse and enhance multi-scale features. It contains top-down and bottom-up paths to achieve the interaction of features at different levels. The third part is the detection head network, which is used to perform the final target detection task.

4. The ground penetrating radar image recognition method based on GDS-YOLOv8n according to claim 1, characterized in that: The steps of using the GDS-YOLOv8n model to identify the constructed ground penetrating radar image data set are: Step 4.1: GPR image preprocessing: After the image is input into the model, the image size is adjusted to 640×640 while maintaining the original aspect ratio; Step 4.2: Target feature extraction in ground penetrating radar image: The preprocessed image is input into the backbone network of the GDS-YOLOv8n model for feature extraction. The C2f and DRRB modules in the backbone network extract multi-scale features from the input image, and the Conv and SPD-Conv modules perform downsampling to generate feature maps of different resolutions. Step 4.3: Enhance feature extraction and feature fusion: In the neck network, feature pyramid network and path aggregation network are used to fuse feature maps of different resolutions; Step 4.4: Target Identification: Three YoloHead detection heads are used to predict the position and category of objects on feature maps with resolutions of 20×20, 40×40, and 80×80.

5. The ground penetrating radar image recognition method based on GDS-YOLOv8n according to claim 1, characterized in that: The specific steps to deploy the GDS-YOLOv8n model to the NVIDIA Jetson edge device are: First, convert the trained .pt model file into a .onnx file on the server; then use the TensorRT of the Jetson Orin NX device to optimize the .onnx model file; finally generate the .engine file.

Images