Asphalt pavement crack detection method and system suitable for mobile platform

Abstract

The invention relates to the technical field of image processing, in particular to an asphalt pavement crack detection method and system suitable for a mobile platform, and the method comprises the steps: obtaining pavement crack images in a public data set and a self-built data set, carrying out the preprocessing optimization and marking of the pavement crack images, and generating a data set; the data set comprises a training set, a verification set and a test set; inputting a training set in the data set into a YOLOv5 network, wherein the YOLOv5 network comprises a trunk feature extraction network, a neck feature fusion network and three detection heads; constructing a road crack detection model based on improvement of a trunk feature extraction network in the YOLOv5 network; inputting a test set in the data set into the road crack detection model for performance evaluation; the complexity of network and operation is greatly reduced, memory access is reduced, the precision is improved compared with that of a source code, and algorithm deployment on embedded hardware equipment is facilitated.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to an asphalt pavement crack detection method and system suitable for a mobile platform. Background Art

[0002] With the improvement of national infrastructure, roads have gradually shifted from large-scale construction to comprehensive maintenance. The increasing mileage of highways has put forward higher requirements for road maintenance. Traditional road crack detection is mainly based on manual detection, which requires certain experience to accurately identify the crack type based on human eye recognition. This method has high deployment cost and low efficiency, and cannot meet the actual needs of existing road maintenance. According to the pain points of manual detection, semi-automatic detection has been developed. Road images are collected by cameras or other imaging devices installed on vehicles, and then image processing technology and image recognition algorithms are used to classify the types of diseases collected. Among them, the operation of image recognition algorithms has high requirements for hardware computing power, and the recognition accuracy is good or bad, and these hardware devices with certain computing power are relatively expensive. Commonly used target detection algorithms include YOLO series, SSD, Faster R-CNN, etc. These algorithms need to meet certain hardware requirements to ensure that the detection accuracy and speed meet the requirements. Improvements to these algorithms can compress the algorithm size and memory requirements without affecting their detection accuracy and time, so as to facilitate the operation of mobile devices.

[0003] Commonly used lightweight networks include mobilenet and shufflenet. These networks are of low complexity and are suitable for running on mobile terminals or other embedded hardware. Commonly used lightweight convolutions include depth-separable convolution operations and partial convolution operations. However, these have low accuracy in detecting road cracks. Too few parameters or insufficient depth of the convolution layer affect the performance of network feature extraction. Other auxiliary modules are also needed to improve accuracy without reducing the number of parameters and algorithm complexity. In addition, the scale of photographic equipment used to shoot cracks is limited, and it is difficult to take high-resolution and clear images. Although most cracks can be observed with the naked eye, there are still a small part that is difficult to distinguish with the human eye. The resulting data set quality is not high, which greatly reduces the feature extraction performance in the target detection network. Achieving a high-precision detection data set for target detection is the key, so that the network can see the "features" clearly and enhance the details of the picture. In this regard, the YOLO series is relatively limited in image enhancement. Improving the quality of the data set on the basis of lightweight can also improve the performance of network detection.

[0004] Therefore, a road damage detection algorithm suitable for mobile terminals with higher lightweight characteristics and higher recognition accuracy is proposed. This algorithm not only needs to improve the accuracy of feature extraction while ensuring the lightweight of the model, but also needs to improve the quality of the data set through image enhancement technology to improve the overall detection performance. Summary of the invention

[0005] The purpose of the present invention is to provide an asphalt pavement crack detection method and system suitable for a mobile platform to solve the problems raised in the above background technology.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] A method for detecting cracks in an asphalt pavement suitable for a mobile platform, the method comprising:

[0008] Obtaining pavement crack images from public data sets and self-built data sets, preprocessing, optimizing and annotating the pavement crack images to generate a data set; the data set includes a training set, a validation set and a test set;

[0009] Input the training set in the data set into the YOLOv5 network, wherein the YOLOv5 network includes a backbone feature extraction network, a neck feature fusion network and three detection heads;

[0010] A road crack detection model is constructed based on the improvement of the backbone feature extraction network in the YOLOv5 network;

[0011] The test set in the dataset is input into the road crack detection model for performance evaluation.

[0012] Preferably, the YOLOv5 network includes a backbone feature extraction network, a neck feature fusion network and three detection heads, including:

[0013] The backbone feature extraction network is used to extract image features of pavement crack images;

[0014] The neck feature fusion network is used to perform feature fusion on the extracted image features;

[0015] The three detection heads are used to predict the bounding box and category according to the image after feature fusion. After the training is completed, the weight file required for the subsequent prediction process is obtained, and the training is completed.

[0016] Preferably, a road crack detection model is constructed based on an improvement of a backbone feature extraction network in a YOLOv5 network, including:

[0017] Enhance the pavement crack image using a generative adversarial network; the generative adversarial network includes a generative model and a discriminative model;

[0018] Based on the bottleneck module of the C3 module in the backbone feature extraction network, the FasterBlock module is constructed;

[0019] Based on the common convolution module in the backbone feature extraction network, the SCDown downsampling module is added;

[0020] Based on the traditional interpolation upsampling module in the YOLOv5 network, it is replaced by the CARAFE upsampling module.

[0021] Preferably, the generative adversarial network includes a generative model and a discriminative model, including:

[0022] The generative model is used to capture the distribution of sample data:

[0023] A1. The generative model receives a low-resolution road crack image as input and extracts basic features of the image through an initial convolutional layer and a batch normalization layer;

[0024] A2, passing the image feature data through a series of residual blocks, wherein the residual block includes two convolutional layers and a batch normalization layer;

[0025] A3. Insert the PReLU activation function and use skip connections to alleviate the gradient vanishing problem and promote the flow of deep information. Upsample through the pixel rearrangement layer to convert the channel information in the feature map into spatial information, achieve efficient and fast upsampling, and convert the feature map into a high-resolution road crack image.

[0026] The discriminant model is used to use a discriminator network to discriminate whether an input high-resolution pavement crack image is real data or a generated sample, and to update the data using a loss function;

[0027] The loss function formula is:

[0028] Where x represents the sample content and a represents the loss parameter.

[0029] Preferably, a FasterBlock module is constructed based on the bottleneck module of the C3 module in the backbone feature extraction network, including:

[0030] The FasterBlock module includes a bottleneck module, a partial convolution module PConv and an EMAttention attention module;

[0031] The bottleneck module is a module in the traditional C3 module in the YOLOv5 network;

[0032] The partial convolution module PConv performs convolution operation on part of the input feature map, and keeps the remaining part unchanged, and then splices the feature map after the convolution operation and the feature map without the convolution operation to form a new feature map, and then uses point convolution to perform convolution operation PWConv on this basis to obtain feature maps of different channels for feature extraction, uses residual links to reuse features, and then redistributes weights;

[0033] The EMAttention attention module is used to improve the feature extraction capability and make up for the deficiencies of some convolution operations in feature extraction.

[0034] Preferably, the EMAttention attention module is used to improve the feature extraction capability and make up for the deficiencies of some convolution operations in feature extraction, including:

[0035] B1. Obtain the feature maps of the different channels as the input feature maps of the EMAttention attention module, divide the input feature maps into G subgroups, each subgroup has C / G channels; then perform average pooling on each subgroup in width W and height H to obtain a feature map with a shape of C / / G×1×W and C / / G×H×1, and then retain the feature map of C / / G×H×W. The three branches perform different operations, and the outputs of the X average pooling branch and the Y average pooling branch are connected along the channel direction to form a new feature map with a shape of C / / G×1×(W+H);

[0036] B2. Apply a 1x1 convolutional layer to the new feature map to integrate the information from the three branches. Use the sigmoid function to perform nonlinear transformation on the connected feature map to obtain two outputs with shapes of C / / G×1×W and C / / G×H×1, and then redistribute the weights. Then perform softmax operation on the feature map after group normalization and global average pooling to obtain probability distribution.

[0037] B3. Perform matrix multiplication on the probability distribution and the input feature map. After the 3×3 convolution operation, the third branch also performs global average pooling and softmax operation. The group normalized feature map and the feature map after softmax processing of the third branch are matrix multiplied. The feature map after softmax processing of the first branch and the feature map after 3×3 convolution operation of the third branch are matrix multiplied to obtain two 1×H×W feature maps. These two feature maps are added and then processed by the sigmoid function to obtain an output feature map. At this time, this output feature map is applied as a weight to the original input feature map to achieve re-weighting of the feature map.

[0038] Preferably, the SCDown downsampling module includes:

[0039] Different from ordinary convolution, the SCDown downsampling module uses a 1×1 convolution kernel with a stride of 1 to adjust the number of channels; and uses a second 3×3 convolution kernel to perform a convolution operation with a stride of 2 and a padding of 1 to achieve lightweight downsampling.

[0040] Preferably, the CARAFE upsampling module comprises: an upsampling kernel prediction module and a feature recombination module;

[0041] The upsampling kernel prediction module includes:

[0042] Input feature map χ with original size C×H×W, take upsampling rate σ as an integer, upsampling kernel size is K up ×K up , first use a 1×1 convolution kernel to adjust the channel C of the input feature map χ to C m To reduce the amount of calculation and parameters, use a size of K encode ×K encode The convolution kernel generates a new upsampling kernel according to the input feature map, and obtains H×W×σ 2 k up 2 The feature map is then expanded in the spatial dimension to obtain a channel dimension of size σH×σW×K up 2 The upsampling kernel is then input into the Softmax function for normalization;

[0043] The feature recombination module comprises:

[0044] For each position of the output feature map, map it to the original input feature map and take out the K-sized up ×K up The area of ​​​​the point is dot-producted with the kernel predicted at that point to obtain an output feature map χ' of size σH×σW×C. Different channels but the same position share the same upsampling kernel.

[0045] Asphalt pavement crack detection system, the system includes: data acquisition module, network input module, model building module and performance evaluation module;

[0046] The data acquisition module is used to obtain pavement crack images in public data sets and self-built data sets, and pre-process, optimize and annotate the pavement crack images to generate a data set; the data set includes a training set, a validation set and a test set;

[0047] Inputting the training set in the data set into the YOLOv5 network through the network input module, wherein the YOLOv5 network includes a backbone feature extraction network, a neck feature fusion network and three detection heads;

[0048] Building a road crack detection model based on the improvement of the backbone feature extraction network in the YOLOv5 network through the model building module;

[0049] The test set in the data set is input into the road crack detection model through the performance evaluation module for performance evaluation.

[0050] Preferably, the model building module includes a generative adversarial network, a FasterBlock module, a SCDown downsampling module and a CARAFE upsampling module;

[0051] The generative adversarial network is used to enhance the pavement crack image;

[0052] The FasterBlock module is used to optimize the image for feature extraction;

[0053] The SCDown downsampling module is used to optimize the image parameters and computational complexity for feature extraction;

[0054] The CARAFE upsampling module is used to replace the traditional interpolation upsampling module in the YOLOv5 network.

[0055] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0056] The present invention optimizes images so that low-definition images can also be well utilized, and at the same time, crack disease features of images can be increased. Compared with the YOLO version suitable for small-size devices, the accuracy is slightly improved and the number of parameters and algorithm complexity are greatly reduced. In the collection of the database, the algorithm is not complicated in engineering implementation, the repeatability is simple, and the data set can be replaced according to needs. The generative adversarial network is used to enhance the characteristics of road diseases in the image, and the low-resolution image is optimized for high resolution. The detection network can adapt to the crack size to obtain clear and prominent crack disease details. Better distinction is made than the original image, which is conducive to algorithm extraction, enhances the robustness of the algorithm, and is suitable for images of different resolutions. The complexity of the network and calculation is greatly reduced, the memory access is reduced, the accuracy is improved compared to the source code, and it is convenient to deploy the algorithm on embedded hardware devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0058] Figure 1 It is a flow chart of a method for detecting cracks in an asphalt pavement applicable to a mobile platform of the present invention;

[0059] Figure 2 It is the network structure diagram of the road crack detection model;

[0060] Figure 3 is a structural diagram of the FasterBlock module of the present invention;

[0061] Figure 4 It is a structural diagram of the EMAttention module of the present invention;

[0062] Figure 5 It is a structural diagram of the SCDown downsampling module of the present invention;

[0063] Figure 6 This is the working principle of the partial convolution module PConv of the present invention;

[0064] Figure 7 This is the working principle of the CARAFE upsampling module of the present invention;

[0065] Figure 8 It is a detection effect diagram of the road crack detection model of the present invention;

[0066] Fig. 9 is a schematic diagram of the pixel recombination of the present invention;

[0067] Fig.10 This is a diagram showing the effect of the crack characteristics being enhanced according to the present invention. DETAILED DESCRIPTION

[0068] 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.

[0069] See also Figure 1-Figure 10 , the present invention provides a technical solution:

[0070] A method for detecting cracks in an asphalt pavement suitable for a mobile platform, the method comprising:

[0071] Obtaining pavement crack images from public data sets and self-built data sets, preprocessing, optimizing and annotating the pavement crack images to generate a data set; the data set includes a training set, a validation set and a test set;

[0072] Obtaining data: Obtain data through public datasets and self-built datasets, and use labelme for data annotation. The self-built dataset is obtained by using a camera to take crack images on the road. Then perform manual annotation, or shoot a video and have the inspectors extract the images frame by frame, annotate them, and complete the dataset. Then divide the dataset into training set, validation set, and test set in a ratio of 8:1:1; the public dataset is part of the dataset in RDD2022.

[0073] The YOLO series of networks uses Mosaic data augmentation technology, which randomly combines four images into a new image. This method can significantly increase the diversity of the data set, allowing the model to perform better when facing complex backgrounds. It can also mix multiple images and labels to create new samples, adjust the color of the image, and perform affine transformations such as scaling, translating, and rotating the image, and use simple image processing methods to expand the data set.

[0074] Input the training set in the data set into the YOLOv5 network, wherein the YOLOv5 network includes a backbone feature extraction network, a neck feature fusion network and three detection heads;

[0075] Preferably, the YOLOv5 network includes a backbone feature extraction network, a neck feature fusion network and three detection heads, including:

[0076] The backbone feature extraction network is used to extract image features of pavement crack images;

[0077] The neck feature fusion network is used to perform feature fusion on the extracted image features;

[0078] The three detection heads are used to predict the bounding box and category according to the image after feature fusion. After the training is completed, the weight file required for the subsequent prediction process is obtained, and the training is completed;

[0079] Model training: A lightweight detection model is deployed using the single-stage YOLO series algorithm as the benchmark model. The data set is input into the improved YOLOv5 network to start training. The YOLO network contains a backbone feature extraction network, a neck feature fusion network, and three detection heads. To ensure lightweight improvement, different YOLO series networks are tested and the corresponding network structure is selected based on the effect. The backbone network is responsible for extracting features, and the detection head is responsible for predicting bounding boxes and categories based on feature maps. After training, the weight file required for the subsequent prediction process is obtained, and the training is completed.

[0080] A road crack detection model is constructed based on the improvement of the backbone feature extraction network in the YOLOv5 network;

[0081] Preferably, a road crack detection model is constructed based on an improvement of a backbone feature extraction network in a YOLOv5 network, including:

[0082] In order to enable the network to further extract effective features and improve the generalization performance of the recognition system, the generative adversarial network is used to enhance the road crack image. The adversarial training ensures the realism and detail richness of the generated image. The proportion of cracks in the image is not high, and strengthening the crack features plays a significant auxiliary role in the improvement of the lightweight network.

[0083] The generative adversarial network includes a generative model and a discriminative model, with a VGG convolutional network as the backbone;

[0084] Preferably, the generative adversarial network includes a generative model and a discriminative model, including:

[0085] The generative model is used to capture the distribution of sample data:

[0086] A1. The generative model receives a low-resolution road crack image as input and extracts basic features of the image through an initial convolutional layer and a batch normalization layer;

[0087] A2, passing the image feature data through a series of residual blocks, wherein the residual block includes two convolutional layers and a batch normalization layer;

[0088] A3. Insert the PReLU activation function and use skip connections to alleviate the gradient vanishing problem and promote the flow of deep information. Upsample through the pixel rearrangement layer to convert the channel information in the feature map into spatial information, achieve efficient and fast upsampling, and convert the feature map into a high-resolution road crack image.

[0089] The discriminant model is used to use a discriminator network to discriminate whether the input high-resolution pavement crack image is real data or a generated sample, and to update the data using a loss function; the discriminant model is generally a binary classifier, which discriminates whether the input is real data or a generated sample to achieve a fake-real effect;

[0090] The loss function formula is:

[0091] Where x represents the sample content and a represents the loss parameter.

[0092] Based on the bottleneck module of the C3 module in the backbone feature extraction network, the FasterBlock module is constructed;

[0093] Preferably, a FasterBlock module is constructed based on the bottleneck module of the C3 module in the backbone feature extraction network, including:

[0094] The FasterBlock module includes a bottleneck module, a partial convolution module PConv and an EMAttention attention module;

[0095] The bottleneck module is a module in the traditional C3 module in the YOLOv5 network;

[0096] The partial convolution module PConv performs convolution operation on a part of the input feature map, and keeps the remaining part unchanged, and then concatenates the feature map after the convolution operation and the feature map without the convolution operation to form a new feature map;

[0097] Due to the reduction of convolution operations, FLOPs are reduced and computational redundancy is achieved, and memory access is also significantly reduced; the feature maps convolved between different channels are highly similar, and these redundant feature maps also have feature information. On this basis, point convolution is used to perform convolution operation PWConv to obtain feature maps of different channels for feature extraction, and residual links are used to reuse features, and then weights are redistributed;

[0098] Partial convolution will limit the extraction of road crack features. Adding the EMA attention mechanism improves the feature extraction capability and makes up for the shortcomings of partial convolution operations in feature extraction. The EMAttention attention module is used to improve the feature extraction capability and make up for the shortcomings of partial convolution operations in feature extraction.

[0099] Preferably, the EMAttention attention module is used to improve the feature extraction capability and make up for the deficiencies of some convolution operations in feature extraction, including:

[0100] B1. Obtain the feature maps of the different channels as the input feature maps of the EMAttention attention module, divide the input feature maps into G subgroups, each subgroup has C / G channels; then perform average pooling on each subgroup in width W and height H to obtain a feature map with a shape of C / / G×1×W and C / / G×H×1, and then retain the feature map of C / / G×H×W. The three branches perform different operations, and the outputs of the X average pooling branch and the Y average pooling branch are connected along the channel direction to form a new feature map with a shape of C / / G×1×(W+H);

[0101] B2. Apply a 1x1 convolutional layer to the new feature map to integrate the information from the three branches. Use the sigmoid function to perform nonlinear transformation on the connected feature map to obtain two outputs with shapes of C / / G×1×W and C / / G×H×1, and then redistribute the weights. Then perform softmax operation on the feature map after group normalization and global average pooling to obtain probability distribution.

[0102] B3, the probability distribution is subjected to matrix multiplication operation with the input feature map, the third branch is subjected to global average pooling and softmax operation after 3×3 convolution operation, the group normalized feature map is subjected to matrix multiplication operation with the feature map after softmax processing of the third branch, the feature map after softmax processing of the first branch is subjected to matrix multiplication operation with the feature map after 3×3 convolution operation of the third branch, and then two 1×H×W feature maps are obtained; these two feature maps are added and then processed by sigmoid function to obtain an output feature map, and at this time, this output feature map is applied as a weight to the original input feature map to achieve re-weighting of the feature map;

[0103] Among them, the EMA attention mechanism adopts a two-dimensional average pooling method, as shown in the formula:

[0104] Where W represents width, H represents height, and Zc represents feature map size.

[0105] Based on the common convolution module in the backbone feature extraction network, the SCDown downsampling module is added;

[0106] Preferably, the SCDown downsampling module includes:

[0107] Different from ordinary convolution, the SCDown downsampling module uses a 1×1 convolution kernel with a stride of 1 to adjust the number of channels; and uses a second 3×3 convolution kernel to perform a convolution operation with a stride of 2 and a padding of 1 to achieve lightweight downsampling.

[0108] Based on the traditional interpolation upsampling module in the YOLOv5 network, it is replaced by the CARAFE upsampling module.

[0109] Preferably, the CARAFE upsampling module comprises: an upsampling kernel prediction module and a feature recombination module;

[0110] The upsampling kernel prediction module includes:

[0111] Input feature map χ with original size C×H×W, take upsampling rate σ as an integer, upsampling kernel size is K up ×K up , first use a 1×1 convolution kernel to adjust the channel C of the input feature map χ to C m To reduce the amount of calculation and parameters, use a size of K encode ×K encode The convolution kernel generates a new upsampling kernel according to the input feature map, and obtains H×W×σ 2 k up 2 The feature map is then expanded in the spatial dimension to obtain a channel dimension of size σH×σW×Kup 2 The upsampling kernel is then input into the Softmax function for normalization;

[0112] The feature recombination module comprises:

[0113] For each position of the output feature map, map it to the original input feature map and take out the K-sized up ×K up The area of ​​​​the point is dot-producted with the kernel predicted at that point to obtain an output feature map χ' of size σH×σW×C. Different channels but the same position share the same upsampling kernel.

[0114] The test set in the dataset is input into the road crack detection model for performance evaluation.

[0115] Asphalt pavement crack detection system, the system includes: data acquisition module, network input module, model building module and performance evaluation module;

[0116] The data acquisition module is used to obtain pavement crack images in public data sets and self-built data sets, and pre-process, optimize and annotate the pavement crack images to generate a data set; the data set includes a training set, a validation set and a test set;

[0117] The training set in the data set is input into the YOLOv5 network through the network input module, and the YOLOv5 network includes a trunk feature extraction network, a neck feature fusion network and three detection heads;

[0118] Building a road crack detection model based on the improvement of the backbone feature extraction network in the YOLOv5 network through the model building module;

[0119] Preferably, the model building module includes a generative adversarial network, a FasterBlock module, a SCDown downsampling module and a CARAFE upsampling module;

[0120] The generative adversarial network is used to enhance the pavement crack image;

[0121] The FasterBlock module is used to optimize the image for feature extraction;

[0122] The SCDown downsampling module is used to optimize the image parameters and computational complexity for feature extraction;

[0123] The CARAFE upsampling module is used to replace the traditional interpolation upsampling module in the YOLOv5 network.

[0124] Inputting the test set in the data set into the road crack detection model through the performance evaluation module to perform performance evaluation;

[0125] Then obtain experimental data on the performance of the product or result:

[0126]

[0127] It can be seen that the new method is higher in precision, recall, mAP50, and mAP50-95 than YOLOv5n, YOLOv8n, and YOLOv10n, and the number of parameters and model size are greatly reduced compared to YOLOv5n, YOLOv8n, and YOLOv10.

[0128] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for detecting cracks in an asphalt pavement suitable for a mobile platform, characterized in that: The method comprises: Obtaining pavement crack images from public data sets and self-built data sets, preprocessing, optimizing and annotating the pavement crack images to generate a data set; the data set includes a training set, a validation set and a test set; Input the training set in the data set into the YOLOv5 network, wherein the YOLOv5 network includes a backbone feature extraction network, a neck feature fusion network and three detection heads; A road crack detection model is constructed based on the improvement of the backbone feature extraction network in the YOLOv5 network; The test set in the dataset is input into the road crack detection model for performance evaluation.

2. The asphalt pavement crack detection method suitable for a mobile platform according to claim 1, characterized in that: The YOLOv5 network includes a backbone feature extraction network, a neck feature fusion network and three detection heads, including: The backbone feature extraction network is used to extract image features of pavement crack images; The neck feature fusion network is used to perform feature fusion on the extracted image features; The three detection heads are used to predict the bounding box and category according to the image after feature fusion. After the training is completed, the weight file required for the subsequent prediction process is obtained, and the training is completed.

3. The asphalt pavement crack detection method applicable to a mobile platform according to claim 1, characterized in that: The road crack detection model is constructed based on the improvement of the backbone feature extraction network in the YOLOv5 network, including: Using a generative adversarial network to enhance the pavement crack image features; the generative adversarial network includes a generative model and a discriminative model; Based on the bottleneck module of the C3 module in the backbone feature extraction network, the FasterBlock module is constructed; Based on the common convolution module in the backbone feature extraction network, the SCDown downsampling module is added; Based on the traditional interpolation upsampling module in the YOLOv5 network, it is replaced by the CARAFE upsampling module.

4. The asphalt pavement crack detection method applicable to a mobile platform according to claim 3, characterized in that: The generative adversarial network includes a generative model and a discriminative model, including: The generative model is used to capture the distribution of sample data: A1. The generative model receives a low-resolution road crack image as input and extracts basic features of the image through an initial convolutional layer and a batch normalization layer; A2, passing the image feature data through a series of residual blocks, wherein the residual block includes two convolutional layers and a batch normalization layer; A3. Insert the PReLU activation function and use skip connections to alleviate the gradient vanishing problem and promote the flow of deep information. Upsample through the pixel rearrangement layer to convert the channel information in the feature map into spatial information, achieve efficient and fast upsampling, and convert the feature map into a high-resolution road crack image. The discriminant model is used to use a discriminator network to discriminate whether an input high-resolution pavement crack image is real data or a generated sample, and to update the data using a loss function; The loss function formula is: Where x represents the sample content and a represents the loss parameter.

5. The asphalt pavement crack detection method applicable to a mobile platform according to claim 3, characterized in that: The bottleneck module based on the C3 module in the backbone feature extraction network constructs a FasterBlock module, including: The FasterBlock module includes a bottleneck module, a partial convolution module PConv and an EMAttention attention module; The bottleneck module is a module in the traditional C3 module in the YOLOv5 network; The partial convolution module PConv performs convolution operation on part of the input feature map, and keeps the remaining part unchanged, and then splices the feature map after the convolution operation and the feature map without the convolution operation to form a new feature map, and then uses point convolution to perform convolution operation PWConv on this basis to obtain feature maps of different channels for feature extraction, uses residual links to reuse features, and then redistributes weights; The EMAttention attention module is used to improve the feature extraction capability and make up for the deficiencies of some convolution operations in feature extraction.

6. The asphalt pavement crack detection method applicable to a mobile platform according to claim 5, characterized in that: The EMAttention attention module is used to improve the feature extraction capability and make up for the shortcomings of some convolution operations in feature extraction, including: B1. Obtain the feature maps of the different channels as the input feature maps of the EMAttention attention module, divide the input feature maps into G subgroups, each subgroup has C / G channels; then perform average pooling on each subgroup in width W and height H to obtain a feature map with a shape of C / / G×1×W and C / / G×H×1, and then retain the feature map of C / / G×H×W. The three branches perform different operations, and the outputs of the X average pooling branch and the Y average pooling branch are connected along the channel direction to form a new feature map with a shape of C / / G×1×(W+H); B2. Apply a 1x1 convolutional layer to the new feature map to integrate the information from the three branches. Use the sigmoid function to perform nonlinear transformation on the connected feature map to obtain two outputs with shapes of C / / G×1×W and C / / G×H×1, and then redistribute the weights. Then perform softmax operation on the feature map after group normalization and global average pooling to obtain probability distribution. B3. Perform matrix multiplication on the probability distribution and the input feature map. After the 3×3 convolution operation, the third branch also performs global average pooling and softmax operation. The group normalized feature map and the feature map after softmax processing of the third branch are matrix multiplied. The feature map after softmax processing of the first branch and the feature map after 3×3 convolution operation of the third branch are matrix multiplied to obtain two 1×H×W feature maps. These two feature maps are added and then processed by the sigmoid function to obtain an output feature map. At this time, this output feature map is applied as a weight to the original input feature map to achieve re-weighting of the feature map.

7. The asphalt pavement crack detection method applicable to a mobile platform according to claim 3, characterized in that: The SCDown downsampling module includes: Different from ordinary convolution, the SCDown downsampling module uses a 1×1 convolution kernel with a stride of 1 to adjust the number of channels; and uses a second 3×3 convolution kernel to perform a convolution operation with a stride of 2 and a padding of 1 to achieve lightweight downsampling.

8. The asphalt pavement crack detection method applicable to a mobile platform according to claim 3, characterized in that: The CARAFE upsampling module includes: an upsampling kernel prediction module and a feature recombination module; The upsampling kernel prediction module includes: Input feature map χ with original size C×H×W, take upsampling rate σ as an integer, upsampling kernel size is K up ×K up , first use a 1×1 convolution kernel to adjust the channel C of the input feature map χ to C m To reduce the amount of calculation and parameters, use a size of K encode ×K encode The convolution kernel generates a new upsampling kernel according to the input feature map, and obtains H×W×σ 2 k up 2 The feature map is then expanded in the spatial dimension to obtain a channel dimension of size σH×σW×K up 2 The upsampling kernel is then input into the Softmax function for normalization; The feature recombination module comprises: For each position of the output feature map, map it to the original input feature map and take out the K-sized up ×K up The area of ​​​​the point is dot-producted with the kernel predicted at that point to obtain an output feature map χ' of size σH×σW×C. Different channels but the same position share the same upsampling kernel.

9. An asphalt pavement crack detection system for implementing an asphalt pavement crack detection method applicable to a mobile platform as claimed in any one of claims 1 to 8, characterized in that: The system comprises: a data acquisition module, a network input module, a model building module and a performance evaluation module; The data acquisition module is used to obtain pavement crack images in public data sets and self-built data sets, and pre-process, optimize and annotate the pavement crack images to generate a data set; the data set includes a training set, a validation set and a test set; Inputting the training set in the data set into the YOLOv5 network through the network input module, wherein the YOLOv5 network includes a backbone feature extraction network, a neck feature fusion network and three detection heads; Building a road crack detection model based on the improvement of the backbone feature extraction network in the YOLOv5 network through the model building module; The test set in the data set is input into the road crack detection model through the performance evaluation module for performance evaluation.

10. The asphalt pavement crack detection system according to claim 6, characterized in that: The model building module includes a generative adversarial network, a FasterBlock module, a SCDown downsampling module and a CARAFE upsampling module; The generative adversarial network is used to enhance the pavement crack image; The FasterBlock module is used to optimize the image for feature extraction; The SCDown downsampling module is used to optimize the image parameters and computational complexity for feature extraction; The CARAFE upsampling module is used to replace the traditional interpolation upsampling module in the YOLOv5 network.

Images