Deep learning-based methods for identifying roadbed defects, electronic devices, and storage media

Abstract

This invention discloses a deep learning-based method, electronic device, and storage medium for identifying roadbed defects. Specifically, it involves constructing a high-quality standard dataset of highway roadbed B-scan images according to the YOLO dataset standard format, and dividing it into training, validation, and test sets; constructing an improved YOLOv8 target detection network, training the YOLOv8 target detection network using the training set, and obtaining a target recognition model; inputting the roadbed defect image to be identified into the target recognition model for identification, and obtaining the identification result. This invention solves the technical problems of low accuracy and slow detection speed in existing technologies for identifying defect images.

Description

Technical Field

[0001] This invention belongs to the field of roadbed detection technology, and in particular relates to a roadbed defect identification method, electronic equipment and storage medium based on deep learning. Background Technology

[0002] In highway construction, insufficient compaction of the roadbed can easily lead to a loose roadbed. After the road is opened to traffic, vehicle loads cause uneven settlement of the roadbed and differential settlement of the pavement base layer, resulting in voids in the roadbed. In areas with poor hydraulic conditions and at the point of leakage in the drainage system, the roadbed often suffers from waterlogging. Water flow can carry away part of the roadbed, forming cavities. Poor drainage during the rainy season can lead to water accumulation in the roadbed. Long-term immersion and softening of the roadbed reduces its shear strength and can cause cracking.

[0003] Road accidents caused by subgrade defects occur frequently, resulting not only in economic losses but also constantly threatening people's property and personal safety. Pre-detection of subgrade damage can improve its performance, extend road lifespan, and reduce costs. Traditional destructive testing methods for subgrades are not only unrepresentative, inefficient, and costly, but also damage roads, failing to meet modern needs for subgrade defect detection. Currently, non-destructive testing methods such as Ground Penetrating Radar (GPR) offer advantages such as being non-destructive, rapid, accurate, and comprehensive, and are increasingly being used for subgrade defect detection. Although GPR equipment technology has advanced to meet the needs of some engineering surveys, current GPR data processing and automated defect analysis technologies still struggle to handle complex underground environments.

[0004] Chinese invention patent application CN202011357009.X discloses a ground-penetrating radar (GPR) method for detecting roadbed defects based on convolutional neural networks. Specifically, it involves acquiring raw image data from both simulated and actual GPR data, preprocessing these data to obtain preprocessed images, and labeling and storing the target information from both the raw and preprocessed images. Next, a PASCAL VOC dataset is constructed using these images and the labeled information, divided into training, validation, and test sets. Then, the initial values ​​of the anchor box parameters are dynamically set based on the aspect ratio of the labeled target bounding boxes. The constructed Cascade R-CNN network is trained and fine-tuned using the training and validation sets to obtain a convolutional network model. The model performance is then evaluated using the test set, achieving accurate and rapid detection of roadbed defects using GPR. However, this method has low accuracy in identifying defect types in roadbed defect images, requires a long inference time, and has a slow detection speed. Summary of the Invention

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

[0006] This invention provides a method for identifying roadbed defects based on deep learning, comprising the following steps:

[0007] S1: Construct a high-quality standard dataset of highway subgrade B-scan images according to the standard format of the YOLO dataset, and divide it into training set, validation set and test set;

[0008] Step S1 is as follows:

[0009] S1.1: Use a ground-penetrating radar system to detect the actual roadbed in different locations, collect ground-penetrating radar images, and convert the collected ground-penetrating radar images into B-scan format;

[0010] S1.2: Construct roadbed models and models for voids, voids, looseness, and cracks respectively. Use GPRmax software to perform forward modeling on the voids, voids, looseness, and cracks to generate ground-penetrating radar B-scan simulation images.

[0011] S1.3: The ground-penetrating radar B-scan simulation images are preprocessed and divided into training, validation and test sets in a ratio of 7:2:1. The images in the training set are annotated using the LabellImg annotation tool according to the YOLO dataset standard format to obtain a high-quality standard dataset of highway subgrade B-scan images.

[0012] S2: Construct an improved YOLOv8 object detection network. The improved YOLOv8 includes a backbone network, a feature fusion network, and a detection head. The C2f module in the backbone network is integrated with a 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 with the SPPELAN module.

[0013] The large kernel convolutional neural network UniRepLK-Net includes the SmaK Block module and the LarK Block module. The LarKBlock module includes the extended repair block, the SE block, the FFN layer, and the BN module.

[0014] UniRepLK-Net consists of four network units. In the first network unit, two 3×3 convolutional blocks with a stride of 2 in the SmaK Block module are used to process the input feature map, resulting in a C-channel feature map. This C-channel feature map is the input data for the second unit. In the second network unit, a 3×3 convolutional block with a stride of 2 in the LarK Block module is used to process the C-channel feature map of the first unit, resulting in a 2C-channel feature map. This 2C-channel feature map is the input data for the third unit. In the third network unit, a 3×3 convolutional block with a stride of 2 in the SmaK Block module is used to process the 2C-channel feature map of the second unit, resulting in a 4C-channel feature map. This 4C-channel feature map is the input data for the fourth unit. In the fourth network unit, a 3×3 convolutional block with a stride of 2 in the LarK Block module is used to process the 4C-channel feature map of the third unit, resulting in an 8C-channel feature map.

[0015] The SPPELAN module includes large kernel convolution and non-dilated convolution;

[0016] The GAM attention mechanism module includes a channel attention submodule and a spatial attention submodule.

[0017] The channel attention submodule performs global max pooling and global average pooling on the channel input feature map with dimensions of width × height × channel, respectively, to obtain a first feature map and a second feature map with dimensions of 1×1× channel. The first and second feature maps are simultaneously input into a multilayer perceptron network. After activation by the ReLU activation function in the multilayer perceptron network, a third and fourth feature map are obtained. The third and fourth feature maps are added to obtain a channel attention weight map with dimensions of 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 with dimensions of width × height × channel to obtain the channel output feature map with dimensions of width × height × channel.

[0018] In the spatial attention submodule, global max pooling and global average pooling are performed on the spatial input feature map with dimensions of width × height × channels, respectively, to obtain a fifth feature map with dimensions of width × height × 1 and a sixth feature map with dimensions of width × height × 1. The fifth and sixth feature maps are concatenated to obtain a seventh feature map. The seventh feature map is input into a 7x7 convolutional block for dimensionality reduction to obtain an eighth feature map. The eighth feature map is input into a multilayer perceptron network and processed by the sigmoid activation function to obtain a ninth feature map. The ninth feature map is multiplied with the spatial input feature map to obtain a spatial output feature map with dimensions of width × height × channels.

[0019] S3: Train the YOLOv8 object detection network using the training set obtained in step 1 to obtain the object recognition model;

[0020] S4: Input the image of the roadbed defects to be identified into the target recognition model for recognition and obtain the recognition result.

[0021] The test set is input into the improved YOLOv8 detection model, which outputs the category and bounding box of highway subgrade defects, thereby realizing the identification of highway subgrade defects.

[0022] 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 run on the processor, and the processor is used to implement the above-mentioned deep learning-based roadbed defect identification method when executing the computer program.

[0023] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described deep learning-based roadbed defect identification method.

[0024] The beneficial effects of this invention are:

[0025] 1) This invention introduces a large kernel convolutional neural network, UniRepLK-Net, into the C2f module to significantly improve recognition accuracy by utilizing large kernel features; a global attention mechanism (GAM) is added to the backbone to enhance semantic and positional information in the features; and a new lightweight network architecture, namely the Efficient Layer Aggregation Network (ELAN), is used in conjunction with SPP to replace the SPPF module, thereby improving the model's capture of target details and enhancing the model's lightweight nature, inference speed, and accuracy.

[0026] 2) This invention utilizes a self-built high-quality standard dataset of clearly defined diseases to conduct model comparison experiments with three algorithms: YOLOv5, YOLOv8, and YOLOv10. For the identification of four diseases—void, crack, looseness, and delamination—the improved YOLOv8 algorithm achieved mean accuracy (mAP@0.5), mean mAP@0.5:0.95, and mean F1 score of 91.3%, 72.3%, and 88.2% respectively on the dataset constructed in this paper. These figures represent improvements of 2.56%, 2.85%, and 2.42% compared to the traditional YOLOv8 algorithm, respectively, demonstrating better detection accuracy and performance. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a flowchart illustrating the roadbed defect identification method based on deep learning of the present invention.

[0029] Figure 2 These are the results of image preprocessing and augmentation in this invention: (a) is the original ground-penetrating radar image and its brightness enhancement result; (b) is the original ground-penetrating radar image and its contrast enhancement result; (c) is the original ground-penetrating radar image and its HSV data enhancement result; (d) is the original ground-penetrating radar image and its noise-added result; (e) is the original ground-penetrating radar image and its flipped result; (f) is the original ground-penetrating radar image and its rotated result; (g) is the original ground-penetrating radar image and its shearing result; (h) is the original ground-penetrating radar image and its translation result; and (i) is the original ground-penetrating radar image and its affine result.

[0030] Figure 3 This is a schematic diagram of the results of labeling the preprocessed image with LabelImg software in this invention, wherein (a) is a schematic diagram of cavity signal labeling, (b) is a schematic diagram of crack signal labeling, (c) is a schematic diagram of loose signal labeling, and (d) is a schematic diagram of void signal labeling.

[0031] Figure 4 This is a schematic diagram of the improved YOLOv8 model network structure of this invention;

[0032] Figure 5 This is a schematic diagram of the UniRepLK-Net structure in this invention;

[0033] Figure 6 This is a schematic diagram of the GAM attention mechanism in this invention;

[0034] Figure 7 This is a schematic diagram of the channel attention submodule structure in this invention;

[0035] Figure 8 This is a schematic diagram of the spatial attention submodule structure in this invention;

[0036] Figure 9 This is a schematic diagram of the ELAN module structure in this invention;

[0037] Figure 10These are the target detection results of roadbed defects using ground-penetrating radar in this invention, where (a) is the image of the roadbed defects to be detected by ground-penetrating radar, (b) is the target detection result of roadbed defects using YOLOv10, (c) is the target detection result of roadbed defects using the improved YOLOv8, (d) is the target detection result of roadbed defects using YOLOv8, and (e) is the target detection result of roadbed defects using YOLOv5. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] S1: Construct a high-quality standard dataset of highway subgrade 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;

[0040] Step S1 is as follows: S1.1: Use a ground-penetrating radar (GPR) system to detect the actual roadbed in different locations, collect GPR images, and convert the collected GPR images into B-scan format; S1.2: Construct roadbed models and models for voids, voids, looseness, and cracks, respectively. Use GPRmax software to perform forward modeling of the void, void, looseness, and crack models to generate GPR B-scan simulation images; S1.3: Preprocess the GPR B-scan simulation images and divide them into training, validation, and test sets in a 7:2:1 ratio. Use the LabellImg annotation tool to annotate the images in the training set according to the YOLO dataset standard format to obtain a high-quality standard dataset of highway roadbed B-scan images. The relevant parameters selected when using GPRmax software for forward modeling are shown in Table 1.

[0041] Table 1 GPRmax forward simulation parameters

[0042]

[0043] In this embodiment, the simulation space size is 6m × 2m × 0.002m. Specifically, the simulation space has a length of 6m along the x-axis, a length of 2m along the y-axis, and a length of 0.002m along the z-axis. Discrete points are taken every 0.002m along the x-axis, y-axis, and z-axis to divide the simulation space. In this embodiment, a duration of 3 × 10⁻⁶ m is selected. -8 The advantage of using a time window of s is to avoid unnecessary computation and improve simulation efficiency. In this embodiment, Ricker wavelet is selected as the excitation source because Ricker wavelet is similar to the pulse wave of ground-penetrating radar electromagnetic wave, and the excitation source frequencies are 400MHz, 800MHz and 1000MHz respectively.

[0044] Figure 2 This is the result of image preprocessing and augmentation in this invention. In this embodiment, the images are subjected to operations such as direct wave removal, noise superposition, brightness enhancement, contrast enhancement, image flipping, rotation, affine transformation, shear transformation, hue-saturation-value (HSV) data enhancement, and translation augmentation to simulate the radar image processing and augmentation process, ultimately obtaining 1260 images of roadbed defects.

[0045] Figure 3 This diagram illustrates the results of labeling the preprocessed image using LabelImg software in this invention. It can be seen that the red rectangle represents the void signal (Empty, E), the green rectangle represents the crack signal (Fissure, F), the blue rectangle represents the loose signal (Loss, L), and the brown rectangle represents the void signal (Void, V).

[0046] S2: Construct a YOLOv8 object detection network. YOLOv8 includes a backbone network, a feature fusion network, and a detection head. The backbone network has an eleven-layer structure, including layers 1, 2, 3, 4, 5, 6, 7, 8, 10, and 11. Layer 1 is a convolutional module, layer 2 is a convolutional module, layer 3 is a C2f_UniRepLK-Net module, layer 4 is a convolutional module, layer 5 is a C2f_UniRepLK-Net module, layer 6 is a convolutional module, layer 7 is a C2f_UniRepLK-Net module, layer 8 is a convolutional module, layer 9 is a C2f_UniRepLK-Net module, and layers 10 and 11 contain a SPPELAN module for GAM (Global Attention Mechanism). This application uses C2f (Cross Stage Partial Feature) in the third layer of the backbone network. The Fusion (cross-stage local feature fusion) module incorporates the large-kernel convolutional neural network UniRepLK-Net. At layers 10 and 11 of the backbone network, a GAM attention mechanism containing a SPPF (Spatial Pyramid Pooling Fast) module is added. The SPPF module in the GAM attention mechanism is replaced with the SPPELAN (Spatial Pyramid Pooling - Efficient Layer Aggregation Network) module. The YOLOv8 object detection network structure constructed in this invention is shown in the diagram below. Figure 4 As shown,

[0047] like Figure 4The diagram shown is the structure of the YOLOv8 object detection network proposed in this application. The network operation flow is as follows: An image with a size of 640×640 and 3 channels is used as the input to the backbone network. The image first enters two consecutive convolutional layers, each with a 3×3 kernel and a stride of 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 on the input image, reducing the amount of data processed subsequently while extracting some basic image features. Then, the image is sequentially connected to the C2f_UniRepLKNetblock module. Inside the C2f_UniRepLKNetblock module, the features are first transformed by a convolutional module, then split into different branches. Each branch then undergoes feature extraction and processing through three UniRepLKNetblock modules. Finally, the outputs of two branches are concatenated to integrate the features. Following the RepLKNetblock module, a Conv convolutional module is connected, also using a 3×3 convolutional kernel with a stride of 2, to continue downsampling and feature extraction of the feature map, further reducing its size. The feature map output from the Conv convolutional module is then fed back into the C2f_UniRepLKNetblock, and this process is repeated multiple times. The feature map processed by multiple C2f_UniRepLKNet modules is then fed into the global attention mechanism module to calculate global attention weights, obtaining the importance weight of each position in the feature map relative to all other positions. The importance weights are then multiplied element-wise with the original feature map to highlight the feature regions that are more critical to the object detection task and suppress relatively unimportant information. The processed feature map is then fed into the Transition layer for dimensionality transformation. Subsequently, the dimensionality-transformed feature map is fed into multiple parallel max pooling layers to extract multi-scale features. The feature maps of different scales are then concatenated and fed into the convolutional layer for fusion and adjustment, completing the backbone network feature extraction.

[0048] Feature fusion network Figure 1 The Neck network, a feature fusion network, first concatenates the feature map output by the backbone network with feature maps from other layers. Then, the concatenated feature map is input into the upsampling layer for amplification. The amplified feature map is then concatenated with the shallow feature map a second time and input into the C2f module. Inside the C2f module, the feature map is processed through multiple 3×3 convolutional layers for feature extraction and fusion.

[0049] like Figure 5As shown, in the UniRepLK-Net large kernel convolutional neural network, a depth-gradient 3×3 convolutional block in the SmaK Block is used to replace the Dilated Reparam Block in the LarK Block. UniRepLK-Net is divided into four units. In the first unit, two 3×3 convolutional blocks with a stride of 2 in the SmaK Block are used to process the input feature map, obtaining a C-channel feature map. This C-channel feature map is the input data for the second unit. In the second unit, a 3×3 convolutional block with a stride of 2 in the LarK Block is used to process the C-channel feature map from the first unit, obtaining a 2C-channel feature map. This 2C-channel feature map is the input data for the third unit. In the third unit, the SmaK Block and the LarK Block are used... The Block combination structure processes the 2C-channel feature map of the second unit through a 3×3 convolutional block with a stride of 2 to obtain the 4C-channel feature map, which serves as the input data for the fourth unit. In the fourth unit, a 3×3 convolutional block with a stride of 2 in LarKBlock processes the 4C-channel feature map of the third unit to obtain the 8C-channel feature map.

[0050] like Figure 6 As shown, the GAM attention mechanism module includes a channel attention submodule and a spatial attention submodule. The input feature map F1 is multiplied with the channel attention map Mc channel by channel to obtain important channel features. The important channel features are then multiplied with the spatial attention map Ms element by element to obtain the features processed by the GAM attention module.

[0051] like Figure 7 As shown, the channel attention submodule uses a 3D permutation method to store information in three dimensions. Then, a two-layer multilayer perceptron (MLP) is used to enhance the connection between channels and space. Specifically, the channel attention submodule performs global max 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. The first and second feature maps are simultaneously input into the MLP network, activated by the ReLU activation function to obtain a third and fourth feature map. The third and fourth feature maps are added together to obtain a channel attention weight map of size 1×1×C. This channel attention weight map is then input into the MLP network, activated by the sigmoid activation function, and multiplied by the channel input feature map of size W×H×C to obtain the channel output feature map of size W×H×C.

[0052] like Figure 8As shown, the spatial attention submodule uses two 7×7 Convs to fuse the spatial information of the input map. To ensure consistency with the channel attention submodule, both modules use the same scaling ratio. Specifically, in the spatial attention submodule, global max pooling and global average pooling are performed on the spatial input feature map to obtain a fifth feature map and a sixth feature map of size W×H×1. The fifth and sixth feature maps are concatenated to obtain a seventh feature map. The seventh feature map is then input into a 7x7 convolutional block for dimensionality reduction to obtain an eighth feature map. The eighth feature map is then input into a multilayer perceptron network and processed by the sigmoid activation function to obtain a ninth feature map. The ninth feature map is multiplied by the spatial input feature map to obtain a spatial output feature map of size W×H×C.

[0053] like Figure 9 As shown, the ELAN module consists of VoVNet (One-Shot Aggregation Network) based on CSPNet (Cross Stage Partial Network), and uses a stack-in-computation-block structure to optimize the gradient length of the entire network. In use, data first enters the Transition module for initial processing, then undergoes a Split operation, splitting it into at least two paths. One path flows directly downwards, while the other path passes through three consecutive convolutional layers. The data from these two paths are then merged in the Concatenate module. Finally, the merged data enters the next Transition module. The Transition module performs operations such as dimensionality transformation and feature integration on the data. The Split operation splits the data, allowing some data to bypass convolutional processing while others undergo multiple convolutional layers, increasing the diversity of feature extraction. The Concatenate module concatenates data from different paths, fusing different features and helping the model learn richer information.

[0054] S3: Train the YOLOv8 object detection network using the training set obtained in step 1 to obtain the object recognition model;

[0055] S4: Input the image of the roadbed defects to be identified into the target recognition model for recognition and obtain the recognition result.

[0056] The test set is input into the improved YOLOv8 detection model, which outputs the category and bounding box of highway subgrade defects, thereby realizing the identification of highway subgrade defects.

[0057] The key features of this invention are: introducing a large kernel convolutional neural network, UniRepLK-Net, into the C2f module to significantly improve recognition accuracy by utilizing large kernel features; adding a global attention mechanism to the backbone part to enhance semantic and positional information in the features; and using SPP in conjunction with a new lightweight network architecture, namely a general efficient layer aggregation network, to replace the SPPF module, thereby improving the model's ability to capture target details while enhancing the model's lightweight nature, inference speed, and accuracy.

[0058] The experimental environment was a computer running Windows 11, with an Intel i5-13500H processor and 32GB of memory. The development tool was PyCharm, the development language was Python 3.11.9, and the deep learning framework was PyTorch 2.0.0. Specific experimental training parameters are shown in Table 2.

[0059] Table 2 Experimental parameters

[0060]

[0061] The dataset was divided into training, validation, and test sets in a ratio of 7:2:1 for experimentation, and compared with five commonly used object detection algorithms. The experimental results are shown in Table 3.

[0062] Table 3 Comparison Results of Comparative Experiments

[0063]

[0064] Note: mAP stands for Mean Average Precision, which refers to the average precision across all classes. It is a commonly used evaluation metric in the field of object detection and is used to measure the performance of object detection algorithms.

[0065] P stands for Precision, which is a metric for evaluating the performance of a classification model. It measures the proportion of samples that are predicted to be positive but are actually positive.

[0066] R stands for Recall, which measures the proportion of positive samples correctly predicted by the model out of all actual positive samples.

[0067] AP stands for Average Precision, an important metric for evaluating the performance of classification models, especially object detection models. It is the area under the Precision-recall curve, which is equivalent to taking the average Precision under different classification and recognition thresholds.

[0068] 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.

[0069] 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 (IoU) is between 0.5 and 0.95.

[0070] mF1 stands for "mean F1 score", which is the average F1 score.

[0071] The YOLOv5, YOLOv8, and YOLOv10 used in Table 3 are all different versions of the same algorithm.

[0072] comprehensive Figure 10 As shown in Table 3, the method of this invention achieves higher recognition accuracy. Compared to the YOLOv8 model, YOLOv10 and YOLOv5 are not sensitive to the detection of hidden defects in highway subgrades, and their overall accuracy is relatively low. However, the method of this invention shows improvements in mF1 value, mAP@0.5 value, and mAP@0.5:0.95 value compared to the other three models. Compared to 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 this invention has higher detection accuracy, better overall performance, and better generalization ability, making it more suitable for detecting multi-faceted hidden defects in highway subgrades. It has the fastest detection speed for images of roadbed defects, meeting the requirements for rapid defect detection and is more suitable for roadbed defect detection tasks.

[0073] This invention also discloses an electronic device, including a memory and a processor. The memory stores a computer program that can run on the processor. The processor executes the computer program to implement the steps of the aforementioned disease identification method. 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 this invention. One or more modules / units can be a series of computer program instruction segments capable of performing specific functions. These instruction segments describe the execution process of the computer program in the computer device. The computer device can be a mobile phone, desktop computer, laptop, handheld computer, cloud server, or other computing devices. The computer device may include, but is not limited to, a processor and a memory. For example, the computer device may also include input / output devices, network access devices, buses, etc. The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of a computer device, connecting various parts of the device through various interfaces and lines. Memory can be used to store computer programs and / or modules. The processor implements computer programs by running or executing the computer programs and / or modules stored in memory, and by accessing data stored in memory. Memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system, at least one application program required for a function (such as sound playback, image playback, etc.), etc.; the data storage area can store data created based on the use of the phone (such as audio data, phonebook, etc.). Furthermore, memory can include high-speed random access memory, and can also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital cards (SD cards), flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices. Modules / units integrated into a computer device, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium.

[0074] Based on this understanding, the present invention can implement all or part of the processes in the methods of the above embodiments, or it can be accomplished by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various methods described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0075] 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, it implements the steps of the above-described disease identification method.

[0076] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0077] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for identifying subgrade diseases based on deep learning, characterized in that, Comprising the following steps: S1: Construct a high-quality standard data set of highway subgrade B-scan image according to the YOLO data set standard format, and divide it into a training set, a validation set and a 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, a large kernel convolutional neural network UniRepLK-Net is integrated into the C2f module of the YOLOV8 backbone network, a GAM attention mechanism module containing an SPPF module is added to the YOLOV8 backbone network, the GAM attention mechanism module includes a channel attention submodule and a spatial attention submodule, and an SPPELAN module is used to replace the SPPF module in the GAM attention mechanism, the SPPELAN module includes a large kernel convolution and a non-dilated convolution; YOLOv8 includes a backbone network, a feature fusion network and a detection head, the backbone network is an eleven-layer network structure, wherein 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 is a GAM containing an SPPELAN module; S3: Train the YOLOv8 target detection network using the training set obtained in step 1 to obtain a target recognition model; S4: Input the image to be recognized into the target recognition model for recognition to obtain a recognition result; Step S1 is specifically: S1.1: Use a ground penetrating radar system to detect actual subgrades in different places, collect ground penetrating radar images, and convert the collected ground penetrating radar images into B-scan form; S1.2: Construct a subgrade model and a cavity, void, loose, crack disease model respectively, and use GPRmax software to generate ground penetrating radar B-scan simulation images by forward modeling of the cavity, void, loose, crack disease model; S1.3: Preprocess the ground penetrating radar B-scan simulation images, and divide them into a training set, a validation set and a test set according to a quantity ratio of 7:2:1, use a Labellmg labeling tool to label the images in the training set according to the YOLO data set standard format, and obtain a high-quality standard data set of highway subgrade B-scan images; Step S4 is specifically: input the test set into the improved YOLOv8 detection model, output the category and boundary box of the highway subgrade disease, and realize highway subgrade disease recognition.

2. A storage medium, characterized by The storage medium has a computer program stored thereon, and the computer program is executed by the processor to implement the steps of the deep learning-based subgrade disease identification method of claim 1.

3. An electronic device, comprising: Comprising a memory and a processor; The memory is used to store a computer program executable on the processor; The processor is configured to implement the deep learning-based roadbed disease identification method of claim 1 when executing the computer program.

Images