Deep learning-based roadbed disease identification method, electronic equipment and storage medium

Abstract

The invention discloses a deep learning-based roadbed disease identification method, electronic equipment and a storage medium, and the method specifically comprises the steps: constructing a highway roadbed B-scan image high-quality standard data set according to a YOLO data set standard format, and dividing the highway roadbed B-scan image high-quality standard data set into a training set, a verification set and a test set; constructing an improved YOLOv8 target detection network, obtaining a training set to train the YOLOv8 target detection network, and obtaining a target recognition model; and inputting the highway subgrade disease image to be recognized into the target recognition model for recognition to obtain a recognition result. According to the invention, the technical problems of low recognition accuracy and low detection speed of the disease image in the prior art are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of roadbed detection, and in particular relates to a roadbed disease identification method, electronic equipment and storage medium based on deep learning. Background Art

[0002] In highway construction, insufficient compaction of the roadbed can easily cause the roadbed to become loose. After the road is opened to traffic, the vehicle load causes uneven settlement of the roadbed and differences in settlement of the pavement base, thus causing the roadbed to become hollow. In areas with poor hydraulic conditions and leaking points in the drainage system, waterlogging diseases often occur in the roadbed, and the water flow sweeps away part of the roadbed, forming cavities. Poor drainage during the rainy season leads to water accumulation in the roadbed, and the roadbed is soaked and softened by water for a long time, reducing its shear strength and causing the roadbed to crack.

[0003] Road accidents caused by roadbed diseases occur from time to time, which not only causes certain economic losses, but also threatens people's property and personal safety at all times. Pre-detection of roadbed damage can improve the performance of the roadbed, extend the service life of the road, and reduce economic costs. Traditional roadbed destructive detection methods are not only poorly representative, inefficient, and have high detection costs, but also cause damage to the road. They can no longer meet the modern needs for roadbed disease detection. At present, non-destructive detection methods such as ground penetrating radar (GPR) have the advantages of non-destructive, rapid, accurate, and comprehensive, and are increasingly being used in the detection of roadbed diseases. Although the technical development of ground penetrating radar equipment has been able to meet the needs of some engineering detection, the current ground penetrating radar data processing and disease automatic analysis technology are still difficult to cope with complex underground environments.

[0004] The Chinese invention with application number CN202011357009.X discloses a method for detecting roadbed defects using ground penetrating radar based on convolutional neural networks, which specifically includes: obtaining the original image data of ground penetrating radar simulation and actual acquisition, preprocessing them to obtain preprocessed images, marking and storing the target information of the original and preprocessed images. Afterwards, the PASCAL VOC dataset is constructed using these images and marking information, which is divided into a training set, a validation set, and a test set. Next, the initial value of the anchor frame parameter is dynamically set according to the aspect ratio of the marked target border, and the training set and validation set are used to fine-tune the constructed Cascade R-CNN network to obtain a convolutional network model. The model performance is then evaluated using the test set to achieve accurate and rapid detection of roadbed defects using ground penetrating radar. However, this method has a low accuracy rate in identifying the type of defects in roadbed defect images, and the inference time required for identification is long and the identification speed is slow. Summary of the invention

[0005] The purpose of the present invention is to provide a roadbed disease identification method, electronic equipment and storage medium based on deep learning, so as to solve the technical problems of low roadbed disease identification accuracy and slow identification speed existing in the prior art.

[0006] The present invention provides a roadbed disease identification method based on deep learning, comprising the following steps: S1: Construct a high-quality standard dataset of highway roadbed B-scan images according to the standard format of the YOLO dataset and divide it into training set, validation set and test set; Step S1 is specifically as follows: S1.1: Use the ground penetrating radar system to detect the actual roadbed at different locations, collect ground penetrating radar images, and convert the collected ground penetrating radar images into B-scan format; S1.2: Construct the roadbed model and the cavity, void, looseness and crack disease models respectively, and use GPRmax software to perform forward simulation on the cavity, void, looseness and crack disease models to generate ground penetrating radar B-scan simulation images; S1.3: Preprocess the ground penetrating radar B-scan simulation images and divide them into training set, validation set and test set in a ratio of 7:2:1. Use LabelImg annotation tool to annotate the images in the training set according to the standard format of YOLO dataset to obtain a high-quality standard dataset of highway subgrade B-scan images.

[0007] S2: Construct an improved YOLOv8 target detection network. The improved YOLOv8 includes the backbone network, feature fusion network and detection head. The C2f module in the backbone network is integrated into the large kernel convolutional neural network UniRepLK-Net. A GAM attention mechanism module containing an SPPF module is added to the backbone network, and the SPPF module in the GAM attention mechanism is replaced by the SPPELAN module. The large kernel convolutional neural network UniRepLK-Net includes SmaK Block module and LarK Block module. The LarKBlock module includes extended repair block, SE block, FFN layer and BN module. UniRepLK-Net is divided into four network units. In the first network unit, two 3×3 convolution blocks with a step size of 2 in the SmaK Block module are used to process the input feature map to obtain the feature map of the C channel, which is the input data of the second unit. In the second network unit, a 3×3 convolution block with a step size of 2 in the LarK Block module is used to process the C channel feature map of the first unit to obtain the feature map of the 2C channel, which is the input data of the third unit. In the third network unit, a 3×3 convolution block with a step size of 2 in the SmaK Block module is used to process the 2C channel feature map of the second unit to obtain the feature map of the 4C channel, which is the input data of the fourth unit. In the fourth network unit, a 3×3 convolution block with a step size of 2 in the LarK Block module is used to process the 4C channel feature map of the third unit to obtain the feature map of the 8C channel.

[0008] The SPPELAN module includes large kernel convolution and non-dilated convolution; The GAM attention mechanism module includes a channel attention submodule and a spatial attention submodule.

[0009] Channel attention submodule, the channel attention submodule performs global maximum pooling and global average pooling on the channel input feature map of size width × height × channel, and obtains a first feature map of size 1 × 1 × channel and a second feature map of size 1 × 1 × channel. The first feature map and the second feature map are simultaneously input into the multilayer perceptron network, and the third feature map and the fourth feature map are obtained after being activated by the ReLU activation function in the multilayer perceptron network. The third feature map and the fourth feature map are added to obtain a channel attention weight map of size 1 × 1 × channel. The channel attention weight map is input into the multilayer perceptron network, activated by the sigmoid activation function, and multiplied by the channel input feature map of size width × height × channel to obtain a channel output feature map of size width × height × channel.

[0010] In the spatial attention submodule, the spatial attention submodule performs global maximum pooling and global average pooling on the spatial input feature map with a size of width × height × channel, respectively, to obtain a fifth feature map with a size of width × height × 1 and a sixth feature map with a size of width × height × 1, concatenate the fifth feature map and the sixth feature map to obtain the seventh feature map, input the seventh feature map into a convolution block with a convolution kernel of 7x7 for dimensionality reduction to obtain the eighth feature map, input the eighth feature map into the multi-layer perceptron network and process it with the sigmoid activation function to obtain the ninth feature map, multiply the ninth feature map and the spatial input feature map to obtain a spatial output feature map with a size of width × height × channel.

[0011] S3: Use the training set obtained in step 1 to train the YOLOv8 target detection network to obtain a target recognition model; S4: Input the highway subgrade disease image to be identified into the target recognition model for identification to obtain the identification result.

[0012] The test set is input into the improved YOLOv8 detection model, which outputs the category and bounding box of highway subgrade diseases to realize highway subgrade disease identification.

[0013] The present invention also provides an electronic device including a memory and a processor, wherein the memory is used to store a computer program that can be run on the processor, and the processor is used to implement the above-mentioned roadbed disease identification method based on deep learning when executing the computer program.

[0014] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned roadbed disease identification method based on deep learning are implemented.

[0015] The beneficial effects of the present invention are: 1) The present invention introduces a large kernel convolutional neural network UniRepLK-Net in the C2f module, and uses large kernel features to significantly improve recognition accuracy; adds a global attention mechanism (GAM) to the backbone part to enhance the semantic information and location information in the features; uses SPP combined with a new lightweight network architecture, namely the universal efficient layer aggregation network (ELAN) to replace the SPPF module, to improve the model's capture of target details, while improving the model's lightweight, inference speed and accuracy.

[0016] 2) The present invention uses a self-built high-quality standard data set with clear diseases to conduct model comparison experiments on the improved YOLOv8 algorithm with three algorithms, YOLOv5, YOLOv8, and YOLOv10. For the identification of four diseases, namely, voids, cracks, looseness, and voids, the average precision mean mAP@0.5 value, mAP@0.5:0.95 value, and mean F1 score of the improved YOLOv8 algorithm on the data set constructed in this paper are 91.3%, 72.3%, and 88.2%, which are 2.56%, 2.85%, and 2.42% higher than those of the traditional YOLOv8 algorithm, respectively, achieving better detection accuracy and effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 It is a flow chart of the roadbed disease identification method based on deep learning of the present invention; Figure 2 : These are the result images of the preprocessing and expansion of the image in the present invention, (a) is the original ground penetrating radar image and its brightness enhancement result image, (b) is the original ground penetrating radar image and its contrast enhancement result image, (c) is the original ground penetrating radar image and its HSV data enhancement result image, (d) is the original ground penetrating radar image and its noise addition result image, (e) is the original ground penetrating radar image and its flip result image, (f) is the original ground penetrating radar image and its rotation result image, (g) is the original ground penetrating radar image and its shearing result image, (h) is the original ground penetrating radar image and its translation result image, (i) is the original ground penetrating radar image and its affine result image; Figure 3 Schematic diagrams of the results of labeling the preprocessed images using the LabelImg software in the present invention, wherein (a) is a schematic diagram of the labeling of void signals, (b) is a schematic diagram of the labeling of crack signals, (c) is a schematic diagram of the labeling of loose signals, and (d) is a schematic diagram of the labeling of void signals; Figure 4 It is a schematic diagram of the network structure of the improved YOLOv8 model of the present invention; Figure 5 It is a schematic diagram of the UniRepLK-Net structure in the present invention; Figure 6 is a schematic diagram of the GAM attention mechanism in the present invention; Figure 7 Schematic diagram of the structure of the channel attention submodule in the present invention; Figure 8 It is a schematic diagram of the structure of the spatial attention submodule in the present invention; Fig. 9 It is a schematic diagram of the structure of the ELAN module in the present invention; Fig.10 It is a diagram of the ground penetrating radar roadbed disease target detection result in the present invention, wherein (a) is the ground penetrating radar roadbed disease image to be detected, (b) is the roadbed disease target detection result of YOLOv10, (c) is the improved YOLOv8 roadbed disease target detection result, (d) is the roadbed disease target detection result of YOLOv8, and (e) is the roadbed disease target detection result of YOLOv5. DETAILED DESCRIPTION

[0019] 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.

[0020] S1: Construct a high-quality standard dataset of highway roadbed B-scan images according to the standard format of the YOLO (You Only Look Once) dataset and divide it into training set, validation set and test set; Step S1 is specifically as follows: S1.1: Use the ground penetrating radar system to detect the actual roadbed in different places, collect ground penetrating radar images, and convert the collected ground penetrating radar images into B-scan form; S1.2: Construct the roadbed model and the void, void, looseness, and crack disease models respectively, and use the GPRmax software to perform forward simulation on the void, void, looseness, and crack disease models to generate ground penetrating radar B-scan simulation images; S1.3: Preprocess the ground penetrating radar B-scan simulation images and divide them into training set, verification set, and test set according to the quantity ratio of 7:2:1. Use the LabelImg annotation tool to annotate the images in the training set according to the standard format of the YOLO data set to obtain a high-quality standard data set of highway roadbed B-scan images. When using the GPRmax software for forward simulation, the relevant parameters are selected as shown in Table 1: Table 1 GPRmax forward modeling parameters

[0021] In this embodiment, the size of the simulation space is 6m×2m×0.002m, which specifically means that the length of the simulation space in the x-axis direction is 6m, the length of the simulation space in the y-axis direction is 2m, and the length of the simulation space in the z-axis direction is 0.002m; in the x-axis direction, a discrete point is taken every 0.002m to divide the simulation space, in the y-axis direction, a discrete point is taken every 0.002m to divide the simulation space, and in the z-axis direction, a discrete point is taken every 0.002m to divide the simulation space. In this embodiment, the selected time length is 3×10 -8 The advantage of selecting a time window of this length is that it avoids unnecessary calculations and improves simulation efficiency. In this embodiment, Ricker wavelet is selected as the excitation source because the Ricker wavelet is similar to the pulse wave of the geological radar electromagnetic wave, and uses 400 MHz, 800 MHz and 1000 MHz as the excitation source frequency respectively.

[0022] Figure 2 It is a result diagram of image preprocessing and expansion in the present invention. In this embodiment, the image is subjected to operations such as removing direct waves, noise superposition, brightness enhancement, contrast enhancement, image flipping, rotation, affine transformation, shear transformation, hue-saturation-value (HSV) data enhancement and translation expansion to simulate the radar image processing and expansion process, and finally 1260 highway roadbed disease images are obtained.

[0023] Figure 3 It is a schematic diagram of the result of labeling the preprocessed image with LabelImg software in the present invention. It can be seen that the red rectangular box represents the void signal (Empty, E), the green rectangular box represents the crack signal (Fissure, F), the blue rectangular box represents the loose signal (Loss, L), and the brown rectangular box represents the void signal (Void, V).

[0024] S2: Construct a YOLOv8 target detection network. YOLOv8 includes a backbone network, a feature fusion network and a detection head. The backbone network is an eleven-layer network structure, including the first layer, the second layer, the third layer, the fourth layer, the fifth layer, the sixth layer, the seventh layer, the eighth layer, the tenth layer and the eleventh layer. The first layer is a convolution module, the second layer is a convolution module, the third layer is a C2f_UniRepLK-Net module, the fourth layer is a convolution module, the fifth layer is a C2f_UniRepLK-Net module, the sixth layer is a convolution module, the seventh layer is a C2f_UniRepLK-Net module, the eighth layer is a convolution module, the ninth layer is a C2f_UniRepLK-Net module, the tenth layer and the eleventh layer are GAM (Global Attention Mechanism) containing a SPPELAN module. This application uses the C2f (Cross Stage Partial Feature Extraction) module in the third layer of the backbone network. The large kernel convolutional neural network UniRepLK-Net is integrated into the Fusion (cross-stage local feature fusion) module, the GAM attention mechanism containing the SPPF (Spatial Pyramid Pooling Fast) module is added to the tenth and eleventh layers of the backbone network, and the SPPF module in the GAM attention mechanism is replaced by the SPPELAN (Spatial Pyramid Pooling - Efficient Layer Aggregation Network, a spatial pyramid pooling structure based on an efficient layer aggregation network) module. The YOLOv8 target detection network structure diagram constructed in the present invention is as follows: Figure 4 As shown, like Figure 4As shown in the figure, it is the YOLOv8 target detection network structure diagram proposed in this application. The operation process of the network is as follows: an image with a size of 640×640 and a number of channels of 3 is used as the input of the backbone network. The image first enters two consecutive convolutional layers. The convolution kernel size of each convolutional block is 3×3 and the step size is 2; after two consecutive convolutional module operations, the feature map size of the image is reduced from 640×640 to 160×160. The main purpose of this step is to perform preliminary feature extraction and downsampling of the input image, while reducing the amount of subsequent processing data, extracting some basic image features; then, the image is connected to the C2f_UniRepLKNetblock module in turn. Inside the C2f_UniRepLKNetblock module, the features are first transformed by a convolution module, and then a split operation is performed to split the features into different branches. After that, each branch is subjected to feature extraction and processing through three UniRepLKNetblocks respectively. Finally, the outputs of two of the branches are spliced ​​to integrate the features. After the RepLKNetblock module, a Conv convolution module is connected, which also uses a 3×3 convolution kernel with a step size of 2. The feature map is further downsampled and feature extracted, so that the size of the feature map is further reduced. The feature map output by the Conv convolution module is input into the C2f_UniRepLKNetblock again, and the execution is repeated many times. The feature map processed by multiple C2f_UniRepLKNet modules is input into the global attention mechanism module to calculate the global attention weight to obtain the importance weight of each position in the feature map relative to all other positions. The importance weight is then multiplied element by element with the original feature map to highlight the feature areas that are more critical to the target detection task and suppress relatively unimportant information. The processed feature map is then input into the Transition layer for dimensional transformation. The dimensional transformed feature map is then input into multiple parallel maximum pooling layers to extract multi-scale features, splice feature maps of different scales, and input into the convolution layer for fusion adjustment to complete the backbone network feature extraction. Feature fusion network Figure 1 The Neck network in the feature fusion network first concatenates the feature map output by the backbone network with the feature maps of other layers, and then inputs the concatenated feature map into the upsampling layer for amplification. The amplified feature map is concatenated with the shallow feature map for a second time and input into the C2f module. Inside the C2f module, the feature map is extracted and fused through a convolution layer with multiple 3×3 convolution kernels. like Figure 5As shown in the figure, in the large kernel convolutional neural network UniRepLK-Net, the deep gradient 3×3 convolution block in SmaK Block is used to replace the Dilated Reparam Block of LarK Block; UniRepLK-Net is divided into four units. In the first unit, two 3×3 convolution blocks with a step size of 2 in SmaK Block are used to process the input feature map to obtain the feature map of the C channel, and the C channel feature map is the input data of the second unit; in the second unit, a 3×3 convolution block with a step size of 2 in LarK Block is used to process the C channel feature map of the first unit to obtain the feature map of the 2C channel, and the 2C channel feature map is the input data of the third unit; in the third unit, SmaK Block and LarK Block combination structure, through the 3×3 convolution block with a stride of 2 in the combination structure, the 2C channel feature map of the second unit is processed to obtain the feature map of the 4C channel, and the 4C channel feature map is the input data of the fourth unit; in the fourth unit, a 3×3 convolution block with a stride of 2 in LarKBlock is used to process the 4C channel feature map of the third unit to obtain the feature map of the 8C channel.

[0025] like Figure 6 As shown in the figure, the GAM attention mechanism module includes a channel attention submodule and a spatial attention submodule. The input feature map F1 is multiplied channel by channel with the channel attention map Mc to obtain important channel features, and the important channel features are multiplied element by element with the spatial attention map Ms to obtain features processed by the GAM attention module.

[0026] like Figure 7 As shown in the figure, the channel attention submodule uses a 3D permutation method to store information in three dimensions; then, a two-layer multi-layer perceptron (MLP) is used to enhance the connection between channels and space; specifically, the channel attention submodule is used to perform global maximum pooling and global average pooling on the channel input feature map of size W×H×C (width×height×channel) to obtain a first feature map of size 1×1×C (1×1×channel) and a second feature map of size 1×1×C, and the first feature map and the second feature map are simultaneously input into the multi-layer perceptron network, and the third feature map and the fourth feature map are activated by the ReLU activation function in the multi-layer perceptron network, and the third feature map and the fourth feature map are added to obtain a channel attention weight map of size 1×1×C, and the channel attention weight map is input into the multi-layer perceptron network and multiplied by the channel input feature map of size W×H×C through the sigmoid activation function to obtain a channel output feature map of size W×H×C.

[0027] like Figure 8As shown in the figure, the spatial attention submodule uses two 7×7 Convs to fuse the spatial information of the input image. In order to ensure consistency with the channel attention submodule, the two modules use the same reduction ratio. Specifically, in the spatial attention submodule, the spatial input feature map is used to perform global maximum pooling and global average pooling respectively to obtain the fifth feature map of size W×H×1 and the sixth feature map of size W×H×1. The fifth feature map and the sixth feature map are concatenated to obtain the seventh feature map. The seventh feature map is input into the convolution block with a convolution kernel of 7x7 for dimensionality reduction to obtain the eighth feature map. The eighth feature map is input into the multi-layer perceptron network and processed by the sigmoid activation function to obtain the ninth feature map. The ninth feature map is multiplied by the spatial input feature map to obtain the spatial output feature map of size W×H×C.

[0028] like Fig. 9 As shown in the figure, the ELAN module is composed of VoVNet (One-Shot Aggregation Network) based on CSPNet (Cross Stage Partial Network), and uses the stack in computation block structure to optimize the gradient length of the entire network. When in use, the data first enters the Transition module for preliminary processing, and then undergoes the Split operation and is split into at least two paths. One path flows directly downward, and the other path is processed by three consecutive convolutional layers in turn. After that, the data of the two paths are merged in the Concatenate module, and finally the merged data enters the next Transition module. The Transition module is used to perform operations such as dimensional transformation and feature integration on the data. The Split operation diverts the data so that part of the data does not undergo convolution processing, while the other part undergoes multi-layer convolution, which increases the diversity of feature extraction. The Concatenate module splices the data of different paths and fuses different features, which helps the model learn richer information.

[0029] S3: Use the training set obtained in step 1 to train the YOLOv8 target detection network to obtain a target recognition model; S4: Input the highway subgrade disease image to be identified into the target recognition model for identification to obtain the identification result.

[0030] The test set is input into the improved YOLOv8 detection model, which outputs the category and bounding box of highway subgrade diseases to realize highway subgrade disease identification.

[0031] The inventive point of the present invention is that: the present invention introduces a large kernel convolutional neural network UniRepLK-Net in the C2f module, and uses the large kernel features to significantly improve the recognition accuracy; adds a global attention mechanism in the Backbone part to enhance the semantic information and position information in the features; uses SPP in combination with a new lightweight network architecture, that is, a universal efficient layer aggregation network to replace the SPPF module, to improve the model's capture of target details, while improving the model's lightweight, reasoning speed and accuracy.

[0032] The experimental environment is run on a computer with Windows 11 operating system, Intel i5-13500H processor and 32GB memory. The development tool is PyCharm, the development language is Python 3.11.9, and the deep learning framework is Pytorch 2.0.0. The specific experimental training parameters are shown in Table 2.

[0033] Table 2 Experimental parameters

[0034] The dataset is divided into training set, validation set and test set in the ratio of 7:2:1 for experiments, and compared with 5 commonly used target detection algorithms. The experimental results are shown in Table 3.

[0035] Table 3 Comparison results of comparative experiments

[0036] Note: mAP stands for Mean Average Precision, which refers to the average accuracy of all categories. It is a commonly used evaluation indicator in the field of target detection and is used to measure the performance of target detection algorithms. P stands for Precision, which is an indicator for evaluating the performance of a classification model. It is used to measure the proportion of samples predicted to be positive that are actually positive. R stands for Recall, which is used to measure the proportion of positive samples correctly predicted by the model to all actual positive samples; AP stands for Average Precision, which is an important indicator for evaluating the performance of classification models, especially target detection models. It is the area under the Precision-recall curve, which is equivalent to taking the average of the Precision under different classification recognition thresholds. mAP@0.5 stands for mean Average Precision at IoU =0.5, which is the average precision when the intersection over union (IoU) is 0.5; mAP@0.5:0.95 stands for mean Average Precision at IoU =0.5:0.95, which is the average precision when the intersection-over-union ratio (IoU) is between 0.5 and 0.95; mF1 stands for "mean F1 - score", which means "average F1 value"; The YOLOv5, YOLOv8, and YOLOv10 used in Table 3 are all different versions of the same algorithm.

[0037] comprehensive Fig.10 It can be seen from Table 3 that the method of the present invention can achieve higher recognition accuracy. Compared with the YOLOv8 model, YOLOv10 and YOLOv5 are not sensitive to the detection of hidden defects of highway subgrade, and the overall accuracy is low. Compared with the other three models, the method of the present invention has certain improvements in mF1 value, mAP@0.5 value, and mAP@0.5:0.95 value. Compared with the YOLOv5, YOLOv8 and YOLOv10 models, the mF1 value increased by 0.34%, 2.08%, and 2.30% respectively; the mAP@0.5 value increased by 1.52%, 2.27%, and 2.12% respectively; and the mAP@0.5:0.95 value increased by 1.25%, 3.38%, and 2.00% respectively. In summary, the method of the present invention has higher detection accuracy, better comprehensive performance and generalization ability, and can better perform the task of detecting multiple hidden defects of highway subgrade. The detection speed of highway roadbed disease images is the fastest, which meets the requirements of rapid disease detection and is more suitable for highway roadbed disease detection tasks.

[0038] The present invention also discloses an electronic device, including a memory and a processor; the memory is used to store a computer program that can be run on the processor; the processor is used to implement the steps of the above-mentioned disease identification method when executing the computer program, the computer program can be divided into one or more modules / units, one or more modules / units are stored in the memory and executed by the processor to complete the present invention, and one or more modules / units can be a series of computer program instruction segments that can complete specific functions, and the instruction segments are used to describe the execution process of the computer program in the computer device. The computer device can be a computing device such as a mobile phone, a desktop computer, a notebook, a handheld computer, and a cloud server. The computer device may include, but is not limited to, a processor and a memory. For example, the computer device may also include input and output devices, network access devices, buses, etc. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general processor can be a microprocessor or any conventional processor, etc. The processor is the control center of the computer device, and uses various interfaces and lines to connect the various parts of the entire computer device; the memory can be used to store computer programs and / or modules, and the processor implements the computer program by running or executing the computer program and / or module stored in the memory, and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application required for a function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created according to the use of the mobile phone (such as audio data, phone book, etc.). In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (SecureDigital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices. If the module / unit integrated in the computer device is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium.

[0039] Based on such understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of each of the above-mentioned methods when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form, etc. Computer-readable media can include: any entity or device that can carry computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal and software distribution medium, etc.

[0040] In addition, the present invention also provides a storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above-mentioned disease identification method are implemented.

[0041] It should be noted that the device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0042] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A roadbed disease identification method based on deep learning, characterized in that: The steps include: S1: Construct a high-quality standard dataset of highway roadbed B-scan images according to the standard format of the YOLO dataset and divide it into training set, validation set and test set; S2: construct an improved YOLOv8 target detection network, the improved YOLOv8 includes a backbone network, a feature fusion network and a detection head, integrates a large kernel convolutional neural network UniRepLK-Net into the C2f module of the YOLOV8 backbone network, adds a GAM attention mechanism module containing an SPPF module to the YOLOV8 backbone network, the GAM attention mechanism module includes a channel attention submodule and a spatial attention submodule, and replaces the SPPF module in the GAM attention mechanism with an SPPELAN module, the SPPELAN module includes a large kernel convolution and a non-expanded convolution; S3: Use the training set obtained in step 1 to train the YOLOv8 target detection network to obtain a target recognition model; S4: Input the highway subgrade disease image to be identified into the target recognition model for identification to obtain the identification result.

2. The roadbed disease identification method based on deep learning according to claim 1 is characterized in that: Step S1 is specifically as follows: S1.1: Use the ground penetrating radar system to detect the actual roadbed at different locations, collect ground penetrating radar images, and convert the collected ground penetrating radar images into B-scan format; S1.2: Construct the roadbed model and the cavity, void, looseness and crack disease models respectively, and use GPRmax software to perform forward simulation on the cavity, void, looseness and crack disease models to generate ground penetrating radar B-scan simulation images; S1.3: Preprocess the ground penetrating radar B-scan simulation images and divide them into training set, validation set and test set in a ratio of 7:2:

1. Use LabelImg annotation tool to annotate the images in the training set according to the standard format of YOLO dataset to obtain a high-quality standard dataset of highway subgrade B-scan images.

3. The roadbed disease identification method based on deep learning according to claim 1 is characterized in that: The large kernel convolutional neural network UniRepLK-Net described in step S2 is specifically: The large kernel convolutional neural network UniRepLK-Net includes SmaK Block module and LarK Block module. The LarKBlock module includes extended repair block, SE block, FFN layer and BN module. The large kernel convolutional neural network UniRepLK-Net is divided into four network units. In the first network unit, two 3×3 convolution blocks with a step size of 2 in the SmaKBlock module are used to process the input feature map to obtain the feature map of the C channel, which is the input data of the second unit; in the second network unit, a 3×3 convolution block with a step size of 2 in the LarK Block module is used to process the C channel feature map of the first unit to obtain the feature map of the 2C channel, which is the input data of the third unit; in the third network unit, a 3×3 convolution block with a step size of 2 in the SmaK Block module is used to process the 2C channel feature map of the second unit to obtain the feature map of the 4C channel, which is the input data of the fourth unit; in the fourth network unit, a 3×3 convolution block with a step size of 2 in the LarK Block module is used to process the 4C channel feature map of the third unit to obtain the feature map of the 8C channel.

4. The roadbed disease identification method based on deep learning according to claim 1 is characterized in that: The channel attention submodule performs global maximum pooling and global average pooling on the channel input feature map with a size of width×height×channel, respectively, to obtain a first feature map with a size of 1×1×channel and a second feature map with a size of 1×1×channel, and inputs the first feature map and the second feature map into the multilayer perceptron network at the same time, and obtains a third feature map and a fourth feature map after being activated by a ReLU activation function in the multilayer perceptron network, and adds the third feature map and the fourth feature map to obtain a channel attention weight map with a size of 1×1×channel, and inputs the channel attention weight map into the multilayer perceptron network, and after being activated by a sigmoid activation function, multiplies it by the channel input feature map with a size of width×height×channel, to obtain a channel output feature map with a size of width×height×channel.

5. The roadbed disease identification method based on deep learning according to claim 1 is characterized in that: The spatial attention submodule performs global maximum pooling and global average pooling on the spatial input feature map with a size of width×height×channel, respectively, to obtain a fifth feature map with a size of width×height×1 and a sixth feature map with a size of width×height×1, concatenates the fifth feature map and the sixth feature map to obtain a seventh feature map, inputs the seventh feature map into a convolution block with a convolution kernel of 7x7 for dimensionality reduction to obtain an eighth feature map, inputs the eighth feature map into a multi-layer perceptron network and processes it with a sigmoid activation function to obtain a ninth feature map, multiplies the ninth feature map and the spatial input feature map to obtain a spatial output feature map with a size of width×height×channel.

6. The roadbed disease identification method based on deep learning according to claim 1 is characterized in that: Step S4 is specifically as follows: inputting the test set into the improved YOLOv8 detection model, outputting the category and bounding box of the highway subgrade disease, and realizing the identification of highway subgrade diseases.

7. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, implements the steps of the roadbed disease identification method based on deep learning as described in any one of claims 1 to 6.

8. An electronic device, characterized in that: including memory and processor; The memory is used to store a computer program executable on the processor; The processor is used to implement the roadbed disease identification method based on deep learning described in any one of claims 1 to 6 when executing the computer program.

Images