Ground penetrating radar dam termite nest detection method and device based on deep learning

Abstract

The invention provides a ground penetrating radar dam termite nest detection method and device based on deep learning, and relates to the field of radar target detection, and the method comprises the steps: carrying out the preprocessing of a ground penetrating radar data set, and obtaining a ground penetrating radar B-Scan image set; performing image expansion on the B-Scan image set of the ground penetrating radar through an improved PGGAN network to obtain an expanded ground penetrating radar image set; marking the ground penetrating radar B-Scan image set and the expanded ground penetrating radar image set to obtain a sample image set; the method comprises the following steps: constructing an improved YOLOv8 model through an original YOLOv8 model, an OfficientViT module, a BRA module and a DGST module; and training the improved YOLOv8 model through the sample image set to obtain a trained improved YOLOv8 model, and detecting the dam termite nest through the trained improved YOLOv8 model. According to the method, by adding the OfficientViT module, the BRA module and the DGST module in the original YOLOv8 model, the detection accuracy and the detection speed of the model on three types of targets can be remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of radar target detection, and particularly to a method and device for detecting termite nests in dikes based on deep learning using ground penetrating radar. Background Art

[0002] The key to eliminating termite hazards in dikes is to find termite nests.

[0003] Existing termite nest detection methods mainly include five categories: manual nest excavation method, shallow seismic reflection method, high-density resistivity method, audio frequency method, and ground penetrating radar method.

[0004] The ground penetrating radar method has obvious technical advantages in aspects such as the accuracy of ant nest positioning, resolution, and detection efficiency, and relatively accurate positioning of ant nests and ant channels has been achieved at home and abroad. However, in existing research, the ant nests in ground penetrating radar images are often identified and determined through manual interpretation. There are interferences such as cavities, tree roots, stones, and soil texture difference interfaces in the dike, and their radar echo images are similar to termite nests. Using the manual interpretation method is very likely to result in misjudgment and missed judgment. At the same time, the manual interpretation method also has deficiencies such as strong subjectivity, low recognition efficiency, and poor generalizability. To solve the above problems, it is necessary to study a method for detecting termite nests in dikes with high accuracy and fast detection speed.

[0005] There are differences in electromagnetic properties such as the dielectric constant between termite nests and the surrounding soil in the dike, and these differences will be reflected in the radar echo image. Currently, the research on the automatic recognition of abnormal objects in ground penetrating radar images mainly focuses on two aspects: the production of target image data sets and target detection and recognition algorithms.

[0006] In terms of the production of target image data sets, due to the existence of a large amount of noise and interference in the radar echo signal, the target signal in the ground penetrating radar disease image is weak, which will affect subsequent target recognition. Therefore, it is first necessary to suppress the interference signal to enhance the target signal. In addition, the number of target images of termite nests, stones, and cavities collected in actual dike detection is limited, making it difficult to meet the training needs of the deep neural network model. Therefore, it is necessary to perform image data augmentation to increase the diversity and quantity of the image data set to improve the generalization performance of the model.

[0007] In terms of target detection and recognition algorithms, due to the strong automatic feature extraction ability and non-linear mapping ability of deep learning, it has been widely used in the automatic detection of ground penetrating radar targets in recent years. Currently, the commonly used deep learning target detection algorithms include two-stage detection algorithms and one-stage detection algorithms. The two-stage target detection method first generates candidate regions that may contain targets, and then further classifies the candidate regions. Its characteristic is high detection accuracy, but slow detection speed. The one-stage target detection algorithm converts target detection into a regression problem and outputs the category and location of the target at one time, with the advantage of fast detection speed. The YOLO (You Only Look Once) series of algorithms are classic one-stage detection algorithms and have been widely applied in ground penetrating radar target recognition. Among them, YOLOv8 is improved on the basis of the success of previous YOLO versions. Aiming at being fast, accurate and easy to use, it shows good performance in various object detection and image classification tasks. However, the types of underground targets of dams are diverse and the structures are complex. When using the YOLOv8 model for target detection, there are still deficiencies in feature extraction and judgment, and the detection of ant nests is not precise enough. Summary of the Invention

[0008] In view of this, the purpose of the present invention is to provide a method and device for detecting termite nests in dams by ground penetrating radar based on deep learning, which is used to solve the problem that when the current YOLOv8 model is used for target detection, there are deficiencies in feature extraction and judgment, and the detection of ant nests is not precise enough.

[0009] The present invention provides a method for detecting termite nests in dams by ground penetrating radar based on deep learning, including the steps of:

[0010] S1: Obtain a ground penetrating radar data set of termite nests, stones and cavities in dams, preprocess the ground penetrating radar data set, and obtain a ground penetrating radar B-Scan image set;

[0011] S2: Construct an improved PGGAN network, and expand the ground penetrating radar B-Scan image set through the improved PGGAN network to obtain an expanded ground penetrating radar image set;

[0012] S3: Label the ground penetrating radar B-Scan image set and the expanded ground penetrating radar image set to obtain a sample image set;

[0013] S4: Construct an improved YOLOv8 model through the original YOLOv8 model, EfficientViT module, BRA module and DGST module;

[0014] S5: Train the improved YOLOv8 model through the sample image set to obtain a trained improved YOLOv8 model, and detect termite nests in dams through the trained improved YOLOv8 model.

[0015] Preferably, step S1 is specifically as follows:

[0016] S11: Perform direct wave removal processing on the ground penetrating radar data to enhance the signal characteristics of the termite nests in the dam, and obtain an enhanced ground penetrating radar image;

[0017] S12: Perform cropping processing on the enhanced ground penetrating radar image to obtain a ground penetrating radar B-Scan image;

[0018] S13: Repeat steps S11 - S12 until all the ground penetrating radar data of the termite nests, stones, and cavities in the dam are traversed, and obtain a set of ground penetrating radar B-Scan images containing three types of targets: termite nests, stones, and cavities.

[0019] Preferably, step S2 is specifically as follows:

[0020] S21: Obtain the original PGGAN network, and incorporate a hybrid local channel attention module into the upsampling module and downsampling module of the original PGGAN network to obtain an improved PGGAN network;

[0021] S22: Input the set of ground penetrating radar B-Scan images into the improved PGGAN network, and obtain the weights of the generator and the first loss value through training;

[0022] S23: Repeat step S22 until the first loss value is less than the first preset value, and obtain a trained generator;

[0023] S24: Generate an augmented set of ground penetrating radar images through the trained generator.

[0024] Preferably, step S3 is specifically as follows:

[0025] Use the labelimg tool to annotate the information of three types of targets: termite nests, stones, and cavities for each ground penetrating radar B-Scan image and each augmented ground penetrating radar image, and obtain a set of sample images;

[0026] The target information includes: target type, center coordinates of the annotation box, length of the annotation box, and width of the annotation box; the set of sample images includes: training image set, validation image set, and test image set.

[0027] Preferably, step S4 is specifically as follows:

[0028] S41: Obtain the original YOLOv8 model, and the original YOLOv8 model includes: original Backbone network, original Neck network, and original Head network;

[0029] S42: Retain the SPPF module in the original Backbone network, construct an improved Backbone network by sequentially connecting the EfficientViT module, the SPPF module, and the BRA module, and use the two EfficientViTModule modules in the EfficientViT module and the BRA module as the output of the improved Backbone network;

[0030] S43: Replace all C2f modules in the original Neck network with DGST modules to obtain an improved Neck network;

[0031] S44: Construct an improved YOLOv8 model through the improved Backbone network, the improved Neck network, and the original Head network.

[0032] Preferably:

[0033] The DGST module includes: a first Conv1×1 module, a GConv3×3 module, a second Conv1×1 module, and a third Conv1×1 module;

[0034] The working process of the DGST module is as follows:

[0035] The original feature X is input into the first Conv1×1 module to obtain feature X1, and feature X1 is divided into feature X2 and feature X3 according to a ratio of 3:1;

[0036] The grouped convolution is performed on feature X3 through the GConv3×3 module, and then the output of the GConv3×3 module is subjected to channel rearrangement operation to obtain feature X4; feature X2 and feature X4 are concatenated to obtain feature X5;

[0037] Two consecutive convolution operations are performed on feature X5 through the second Conv1×1 module and the third Conv1×1 module to obtain feature X6; feature X5 and feature X6 are concatenated to obtain the output feature of the DGST module.

[0038] Preferably, step S5 is specifically as follows:

[0039] S51: Input the sample image into the improved YOLOv8 model to obtain the predicted detection result and the second loss value;

[0040] S52: Perform backpropagation on the improved YOLOv8 model through the predicted detection result and the real detection result to adjust the weights of the improved YOLOv8 model;

[0041] S53: Repeat steps S51 - S52 until the second loss value is less than the second preset value to obtain the trained improved YOLOv8 model;

[0042] S54: Input the ground penetrating radar image to be measured into the trained improved YOLOv8 model to obtain the detection result of the termite nest in the dam.

[0043] A ground penetrating radar termite nest detection device based on deep learning, comprising: a processor and a storage medium; the processor loads and executes the instructions and data in the storage medium to implement the above-mentioned ground penetrating radar termite nest detection method based on deep learning.

[0044] The present invention has the following beneficial effects:

[0045] 1. For the original ground penetrating radar data set, the direct wave removal processing is used to enhance the signal characteristics of the termite nest in the dam, which can significantly improve the reflection intensity of the target in the image and provide high-quality samples for the production of the data set in the subsequent target recognition task;

[0046] 2. By adding a hybrid local channel attention module to the original PGGAN network, the quality of the generated image is improved, the limited ground penetrating radar images are expanded, and the generalization ability of subsequent recognition of termite nests, stones and cavities is improved;

[0047] 3. By adding an EfficientViT module, a BRA module and a DGST module to the original YOLOv8 model, the detection accuracy and detection speed of the model for three types of targets can be significantly improved. Description of the Drawings

[0048] Figure 1 It is a flowchart of the method in the embodiment of the present invention;

[0049] Figure 2 It is a comparison diagram of the ground penetrating radar data before and after preprocessing;

[0050] Figure 3 It is a structural diagram of the improved PGGAN network;

[0051] Figure 4 It is an expanded ground penetrating radar image;

[0052] Figure 5 It is a schematic diagram of the annotation of the termite nest target information;

[0053] Figure 6 It is a structural diagram of the improved YOLOv8 model;

[0054] Figure 7 It is a schematic diagram of the detection result of the improved YOLOv8 model on the test set;

[0055] Figure 8 It is a structural diagram of the device in the embodiment of the present invention;

[0056] The implementation, functional features and advantages of the present invention will be further described with reference to the embodiments and the accompanying drawings. Detailed implementation manners

[0057] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0058] Refer to Figure 1 , the present invention provides a ground penetrating radar dam termite nest detection method based on deep learning, including the steps of:

[0059] S1: Obtain a ground penetrating radar data set of dam termite nests, stones and cavities, preprocess the ground penetrating radar data set, and obtain a ground penetrating radar B-Scan image set;

[0060] As an embodiment:

[0061] Step S1 is specifically:

[0062] S11: Perform direct wave removal processing on the ground penetrating radar data to enhance the signal characteristics of the dam termite nest, and obtain an enhanced ground penetrating radar image;

[0063] S12: Perform cropping processing on the enhanced ground penetrating radar image to obtain a ground penetrating radar B-Scan image;

[0064] S13: Repeat steps S11-S12 until all the ground penetrating radar data of the dam termite nests, stones and cavities are traversed, and obtain a ground penetrating radar B-Scan image set containing three types of targets: termite nests, stones and cavities.

[0065] Specifically, use the ground penetrating radar to collect data on-site at different levels of dikes, at different times, and under different weather conditions, in order to record the ground penetrating radar image characteristics of termite nests in different regions and different soil conditions, and preprocess the images. The comparison of the ground penetrating radar data before and after preprocessing is as Figure 2 .

[0066] S2: Construct an improved PGGAN network, and perform image augmentation on the ground penetrating radar B-Scan image set through the improved PGGAN network to obtain an augmented ground penetrating radar image set;

[0067] As an embodiment, the structure of the improved PGGAN network is as Figure 3 shown;

[0068] Step S2 is specifically:

[0069] S21: Obtain the original PGGAN network, and incorporate a hybrid local channel attention module into the upsampling module and the downsampling module of the original PGGAN network to obtain an improved PGGAN network;

[0070] Specifically, an improved PGGAN network is constructed using a Mixed Local Channel Attention (MLCA) module. The MLCA module first undergoes Local Average Pooling (LAP) and Global Average Pooling (GAP) processes. LAP focuses on the features of the local region, while GAP captures the statistical information of the entire feature map; the features after local pooling and global pooling are both subjected to feature transformation through 1D convolution (Conv1d) to compress the feature channels while keeping the spatial dimension unchanged; the features after 1D convolution processing are rearranged (Reshape) to adapt to subsequent operations; for the features after local pooling, after 1D convolution and rearrangement, they are combined with the original input features through a "multiplication" operation (X) to highlight useful features; for the features after global pooling, after 1D convolution and rearrangement, they are combined with the local pooling features through an "addition" operation to achieve the fusion of global context information; finally, the feature map processed by local and global attention is restored to the original spatial dimension through an Unpooling (UNAP) operation.

[0071] S22: Input the ground penetrating radar B-Scan image set into the improved PGGAN network, and obtain the weights of the generator and the first loss value through training;

[0072] Specifically, the network structure parameters of the generator are shown in Table 1, and the network structure parameters of the corresponding discriminator of the generator are shown in Table 2;

[0073] Table 1 Network Structure Parameters of the Generator

[0074]

[0075]

[0076] Table 2 Network Structure Parameters of the Discriminator

[0077]

[0078]

[0079] S23: Repeat step S22 until the first loss value is less than the first preset value to obtain a trained generator;

[0080] S24: Generate an augmented ground penetrating radar image set through the trained generator.

[0081] Specifically, train for 800 rounds, use the saved weights to generate 800 images, and some of the images in the generated augmented ground penetrating radar image set are as Figure 4 shown, among which 562 images containing target features are saved.

[0082] S3: Label the ground penetrating radar B-Scan image set and the augmented ground penetrating radar image set to obtain a sample image set;

[0083] As an example:

[0084] Step S3 is specifically as follows:

[0085] Use the labelimg tool to label the three types of target information, namely termite nests, stones, and cavities, for each ground penetrating radar B-Scan image and each augmented ground penetrating radar image to obtain a sample image set;

[0086] The target information includes: target type, center coordinates of the annotation box, length of the annotation box, and width of the annotation box; The sample image set includes: training image set, validation image set, and test image set.

[0087] Specifically, the labeled information is as Figure 5 shown. Filter out the images containing target features, and use the PGGAN network to train the model and perform data augmentation operations. Finally, 882 high-quality images are retained. The preset ratio can be set to 7:2:1. Divide the target image dataset into a training set, a test set, and a validation set according to the ratio of 7:2:1.

[0088] S4: Construct an improved YOLOv8 model through the original YOLOv8 model, EfficientViT module, BRA module, and DGST module;

[0089] As an example, the structure of the improved YOLOv8 model is as Figure 6 shown;

[0090] Step S4 is specifically as follows:

[0091] S41: Obtain the original YOLOv8 model, which includes: the original Backbone network, the original Neck network, and the original Head network;

[0092] S42: Retain the SPPF module in the original Backbone network, and construct an improved Backbone network by sequentially connecting the EfficientViT module, the SPPF module, and the BRA module. Use the two EfficientViTModule modules in the EfficientViT module and the BRA module as the output of the improved Backbone network;

[0093] Specifically, input the sample images into the Conv convolutional layer, DSConv module, EfficientViT Module module, and SPPF module in sequence to generate ground penetrating radar image feature maps of termite nests, stones, and cavities with different scales and weights;

[0094] Introduce a double - layer routing attention mechanism between the Backbone and the neck. The double - layer routing attention mechanism includes two attention layers, as shown by P2 and P3 in Figure 6 . The first attention layer is used for adaptive feature fusion of the input feature map to extract richer semantic information. The second attention layer is used for weighting the fused features to highlight important target regions;

[0095] S43: Replace all C2f modules in the original Neck network with DGST modules to obtain an improved Neck network;

[0096] Specifically, replace all C2f modules in the Neck with dynamic group shuffle transformers (DGST) modules. The core of this module is a 3:1 partitioning strategy, where a part undergoes group convolution and channel shuffle operations. The convolution operation replaces the linear transformation of the fully - connected layer. This design not only reduces computational requirements but also better adapts to the characteristics of convolutional neural networks, further enhancing the performance of feature fusion in the neck network.

[0097] S44: Construct an improved YOLOv8 model by improving the Backbone network, the improved Neck network, and the original Head network.

[0098] As an example:

[0099] The DGST module includes: the first Conv1×1 module, the GConv3×3 module, the second Conv1×1 module, and the third Conv1×1 module;

[0100] The working process of the DGST module is as follows:

[0101] The original feature X is input into the first Conv1×1 module to obtain feature X1. Feature X1 is divided into feature X2 and feature X3 according to a 3:1 ratio;

[0102] The GConv3×3 module performs grouped convolution on feature X3, and then the output of the GConv3×3 module undergoes a channel rearrangement operation to obtain feature X4. Feature X2 and feature X4 are concatenated to obtain feature X5;

[0103] The second Conv1×1 module and the third Conv1×1 module perform two consecutive convolution operations on feature X5 to obtain feature X6. Feature X5 and feature X6 are concatenated to obtain the output feature of the DGST module.

[0104] S5: Train the improved YOLOv8 model with the sample image set to obtain the trained improved YOLOv8 model, and detect the termite nests on the dam through the trained improved YOLOv8 model.

[0105] As an embodiment:

[0106] Step S5 is specifically as follows:

[0107] S51: Input the sample image into the improved YOLOv8 model to obtain the predicted detection result and the second loss value;

[0108] S52: Perform backpropagation on the improved YOLOv8 model through the predicted detection result and the true detection result to adjust the weights of the improved YOLOv8 model;

[0109] S53: Repeat steps S51 - S52 until the second loss value is less than the second preset value to obtain the trained improved YOLOv8 model;

[0110] S54: Input the ground - penetrating radar image to be tested into the trained improved YOLOv8 model to obtain the detection result of the termite nests on the dam.

[0111] Specifically, the detection performance indicators include: Precision (P), Recall (R), Average Precision (AP), mean Average Precision (mAP), and Frames Per Second (FPS), as follows:

[0112]

[0113] Among them, TP represents the number of negative samples misidentified as positive samples, FP represents the number of samples identified as positive samples, and FN represents the number of positive samples misidentified as negative samples.

[0114] Specifically, the hardware resources: NVIDIA GeForcce RTX3060, Windows10 operating system. The programming language is Python, CUDA is 11.5, the deep - learning framework is Pytorch 1.11.0, and other relevant dependency libraries are installed according to the requirements of the YOLOv8requirement.txt file.

[0115] Parameter settings: The training parameters of the improved YOLOV8 model are as follows. The pre - trained weights are the yolov8s.pt model, the number of training epochs is 200, 16 pictures are input for each training, and the picture size is 640×640.

[0116] Result analysis: Figure 7 Figure 2 shows the schematic diagram of the detection results of the improved YOLOv8 model on the test set, and Table 3 shows the detection performance index results of the improved YOLOV8 model;

[0117] Table 3 Detection performance index results of the improved YOLOV8 model

[0118]

[0119] From Figure 7 and Table 3, it can be seen that the improved index results are generally better than the results before improvement. The precision, mean average precision, and detection speed of the improved YOLOv8 model have all increased to a certain extent. Although the recall rate of the model has decreased to a certain extent, its value is still above 80%, within the minimum recall rate range required in this example. In addition, this detection method mainly focuses on the improvement of mean average precision and detection speed.

[0120] Please refer to Figure 8 , Figure 8 which is the schematic diagram of the hardware device working in the embodiment of the present invention. The hardware device specifically includes: a ground penetrating radar dam termite nest detection device 401 based on deep learning, a processor 402, and a storage medium 403.

[0121] A ground penetrating radar dam termite nest detection device 401 based on deep learning: The ground penetrating radar dam termite nest detection device 401 based on deep learning implements the ground penetrating radar dam termite nest detection method based on deep learning.

[0122] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the ground penetrating radar dam termite nest detection method based on deep learning.

[0123] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the ground penetrating radar dam termite nest detection method based on deep learning.

[0124] It should be noted that in this article, the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or further includes elements inherent to this process, method, article or system. Without more limitations, an element defined by the statement "including a..." does not exclude the existence of another identical element in the process, method, article or system including that element.

[0125] The serial numbers of the embodiments of the present invention above are only for description and do not represent the superiority or inferiority of the embodiments. Among the several unit claims of the devices, several of these devices may be embodied by the same hardware item. The use of the terms first, second, and third does not indicate any order, and these terms may be interpreted as identifiers.

[0126] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A method for detecting termite nests in dikes using ground penetrating radar based on deep learning, characterized in that, Including the steps: S1: Obtain the ground penetrating radar dataset of dike termite nests, stones and cavities, preprocess the ground penetrating radar dataset to obtain the ground penetrating radar B-Scan image set; S2: Construct an improved PGGAN network, and expand the ground penetrating radar B-Scan image set through the improved PGGAN network to obtain an expanded ground penetrating radar image set; S3: Label the ground penetrating radar B-Scan image set and the expanded ground penetrating radar image set to obtain a sample image set; S4: Construct an improved YOLOv8 model through the original YOLOv8 model, EfficientViT module, BRA module and DGST module; S5: Train the improved YOLOv8 model with the sample image set to obtain a trained improved YOLOv8 model, and detect the dike termite nests with the trained improved YOLOv8 model.

2. The ground penetrating radar dam termite nest detection method based on deep learning according to claim 1, characterized in that, The specific steps of S1 are: S11: Perform direct wave removal processing on the ground penetrating radar data to enhance the signal characteristics of the dike termite nests, and obtain an enhanced ground penetrating radar image; S12: Crop the enhanced ground penetrating radar image to obtain a ground penetrating radar B-Scan image; S13: Repeat steps S11 - S12 until all the ground penetrating radar data of dike termite nests, stones and cavities are traversed, and obtain a ground penetrating radar B-Scan image set containing three types of targets: termite nests, stones and cavities.

3. The ground penetrating radar dam termite nest detection method based on deep learning according to claim 1, characterized in that, The specific steps of S2 are: S21: Obtain the original PGGAN network, and incorporate a hybrid local channel attention module into the upsampling module and downsampling module of the original PGGAN network to obtain an improved PGGAN network; S22: Input the ground penetrating radar B-Scan image set into the improved PGGAN network, and obtain the weights of the generator and the first loss value after training; S23: Repeat step S22 until the first loss value is less than the first preset value to obtain a trained generator; S24: Generate an expanded ground penetrating radar image set through the trained generator.

4. The ground penetrating radar dam termite nest detection method based on deep learning according to claim 1, characterized in that The specific steps of S3 are: Use the labelimg tool to label the information of three types of targets: termite nests, stones and cavities for each ground penetrating radar B-Scan image and each expanded ground penetrating radar image to obtain a sample image set; The target information includes: target type, center coordinates of the annotation box, length of the annotation box, and width of the annotation box; the sample image set includes: training image set, validation image set and test image set.

5. The ground penetrating radar dam termite nest detection method based on deep learning according to claim 1, characterized in that, The specific steps of S4 are: S41: Obtain the original YOLOv8 model, and the original YOLOv8 model includes: original Backbone network, original Neck network and original Head network; S42: Retain the SPPF module in the original Backbone network, construct an improved Backbone network by sequentially connecting the EfficientViT module, SPPF module and BRA module, and use the two EfficientViTModule modules in the EfficientViT module and the BRA module as the output of the improved Backbone network; S43: Replace all C2f modules in the original Neck network with DGST modules to obtain an improved Neck network; S44: Construct an improved YOLOv8 model by using the improved Backbone network, the improved Neck network, and the original Head network.

6. The ground penetrating radar dam termite nest detection method based on deep learning according to claim 1, characterized in that: The DGST module includes: a first Conv1×1 module, a GConv3×3 module, a second Conv1×1 module, and a third Conv1×1 module; The working process of the DGST module is as follows: The original feature X is input into the first Conv1×1 module to obtain feature X1, and feature X1 is divided into feature X2 and feature X3 according to a ratio of 3:1; The grouped convolution is performed on feature X3 through the GConv3×3 module, and then the output of the GConv3×3 module is subjected to a channel rearrangement operation to obtain feature X4; feature X2 and feature X4 are concatenated to obtain feature X5; Two consecutive convolution operations are performed on feature X5 through the second Conv1×1 module and the third Conv1×1 module to obtain feature X6; feature X5 and feature X6 are concatenated to obtain the output feature of the DGST module.

7. The ground penetrating radar termite nest detection method based on deep learning according to claim 1, characterized in that, Step S5 is specifically as follows: S51: Input the sample image into the improved YOLOv8 model to obtain a predicted detection result and a second loss value; S52: Perform backpropagation on the improved YOLOv8 model through the predicted detection result and the true detection result to adjust the weights of the improved YOLOv8 model; S53: Repeat steps S51 - S52 until the second loss value is less than the second preset value to obtain a trained improved YOLOv8 model; S54: Input the ground penetrating radar image to be measured into the trained improved YOLOv8 model to obtain the detection result of the dam termite nest.

8. A ground penetrating radar dam termite nest detection device based on deep learning, characterized in that: Including: A processor and a storage medium; the processor loads and executes the instructions and data in the storage medium to implement the ground penetrating radar dam termite nest detection method based on deep learning according to any one of claims 1 to 7.

Images