Research On Highway Pavement Disease Detection Method Based On Deep Learning

In recent years,the rapid development of road construction scale has gradually formed the huge national road network.However,with the daily use of highway,it leads to the formation of a various highway surface diseases,which brings potential safety hazards.At present,the state is actively increasing the investment of highway maintenance to ensure the safety of travel.The detection of road surface disease is the premise of highway maintenance,but the traditional detection methods have the problems of high cost and low accuracy.Therefore,it is very important to realize accurate and efficient road surface disease detection.This thesis studies the detection method of highway pavement disease,combines with deep learning technology to achieve efficient and accurate detection of highway pavement disease,and provides a reference scheme for highway maintenance.The main research contents of this paper are as follows:(1)Firstly,this study collects a number of road surface disease data to make a data set.This thesis collects samples of four types of highway pavement diseases:transverse crack,longitudinal crack,mesh crack and pothole.The data samples are preprocessed,and the space transform,noise perturbation and brightness transform are used to expand the data set.(2)Secondly,according to the problems of low accuracy of target recognition of highway pavement disease.This study proposes a highway pavement disease detection method based on improved YOLOv5 by combining the target detection algorithm.Firstly,adds a scale detection layer,and the appropriate prior box is obtained by Kmeans++ algorithm clustering.Secondly,in order to improve the ability to extract and pay attention to the characteristics of highway surface disease.combines with P＿CBAM attention mechanism to improve the path aggregation network.Finally,ICIoU frame regression loss function is used to replace the original frame regression loss function CIoU,so as to improve the positioning accuracy of the test box.The model YOLOv5-PI used for highway pavement disease detection improves by the above method,and analyzes comparative experimental.,The experimental results show that compared with the original YOLOv5 model,the mAP value of pavement damage detection by the YOLOv5-PI model increase by 3% to 93.4%.(3)Finally,in order to obtain more intuitive and accurate characteristics of road surface disease,combined with semantic segmentation algorithm,a road surface disease segmentation method based on improved DeepLabv3+ is proposed.Firstly,the principle and structure of DeepLabv3+ model are analyzed,and selects the network structure of DCNN module.Then,ResNet101 is determined as the network structure of DCNN module through comparative experiments.Secondly,improves the decoding terminal structure of DeepLapv3+ model is to improve the segmentation accuracy of highway pavement diseases.Finally,compares the analysis experimen.And the experimental results shows that compares with the traditional DeepLabv3+ model,the MIoU value performance of the improved DeepLabv3+ model in the pavement disease data set increased from 65.83% to66.16%.And compares with several mainstream segmentation models,it has better segmentation effect.