Research On Pavement Crack Recognition Method Based On Deep Learning

In recent years,with the continuous growth of China’s economy,infrastructure projects have also undergone rapid development,and highway construction is an important component of the field of infrastructure construction.In highway construction,the issue of post maintenance of roads inevitably arises.Road surface damage is one of the key issues that need to be paid attention to in road maintenance work,and road surface cracks are the most common type of damage.So this study takes road cracks as the research object,using semantic segmentation algorithms as the main tool to identify them,and analyzes and studies the preprocessing of road crack images,expansion of datasets,improvement of classic UNet network models,and calculation of crack length and width.First,according to the characteristics of pavement cracks,we will use the method of combining Gaussian Bilateral filter and minimum filtering algorithm to filter the obtained crack image to eliminate noise interference in the image.Next,we use elastic deformation techniques,including rotation,mirroring,cropping,and scaling,to enhance the data of the original image and corresponding labeled images,in order to solve the problem of insufficient sample data and complete the expansion of the data samples.Meanwhile,in order to evaluate the generalization ability of the model,we divided the test dataset into three categories: the first category is a dataset formed by the expansion and fusion of images collected by Crack500 and recognition vehicles,the second category is a crack dataset expanded by Crack500,and the third category is a crack dataset collected and expanded by recognition vehicles.In order to test the optimal depth of UNet networks,extensive network architecture searches or inefficient integration testing are required.Jumping connections impose unnecessary limitations on the fusion scheme,forcing fusion only when the encoder and decoder subnets have the same proportion of feature maps,which has limitations in road crack segmentation and recognition.To solve this problem,someone has proposed the UNet++network,which borrows the idea of Dense Net and integrates different sizes of UNet structures in the network.Through short connections and upsampling and downsampling operations,multiple features at different levels are simply fused.However,the drawbacks of this method are also evident,as the network is very complex,with a large number of parameters and a significant decrease in speed.This article proposes using the Mobile Net V3 network to replace the encoder part of UNet for feature extraction,thereby improving both performance and speed.The improved UNet Mobile Net V3 network model reduces the number of parameters compared to the classic UNet model,resulting in shorter algorithm runtime and more optimized structure.After experimental verification,the UNet Mobile Net V3 network model performs well on the Crack500 dataset and the crack fusion dataset collected by the recognition vehicle,with a pixel accuracy of 95.32%,an F1 value of 94.4%,and an average intersection to union ratio(m Io U)of 80.35%.In addition,the improved UNet Mobile Net V3 model performs better than traditional UNet models in identifying the remaining two datasets.Moreover,the parameters of the UNet Mobile Net V3 network model are only 13.1M,which is 58% less than the parameters of the traditional road recognition network model UNet.Finally,this article uses the topology refinement algorithm to extract the skeleton of cracks,which can calculate the length of road cracks.Then,based on the boundary tracking method,the area of cracks is calculated,and the width of cracks is finally obtained,achieving quantitative analysis of cracks.At the same time,under the conditions of computers,it is possible to receive large-scale input of images,and the average recognition time of a single image is only 146 milliseconds.This technology can to some extent improve the work efficiency of road maintenance departments,especially in road recognition tasks.Overall,this study successfully completed the task of identifying road cracks.By improving the traditional UNet network model,the identification of cracks was successfully completed,and the calculation of crack width was achieved,providing new ideas for road maintenance and upkeep.In addition,this also provides powerful auxiliary data for relevant decision-making.