Research On Automated Pavement Distress Detection Using Deep Learning

Pavement distress is an important factor that threatens road safety.Accurately and timely identification of pavement distress is critical to ensuring the operation of the safety.Nowadays,vehicular-based distress detection methods has gradually replaced manual investigate methods.Traditional image processing methods are not satisfactory in accuracy and speed.At present,deep learning image recognition technology has been applied to pavement crack detection.However,it is still unable to solve the problem of multi-target detection for different types of distress.And the accuracy of pavement crack identification needs to be improved.Based on the two-dimensional laser image data,this paper proposes a set of automatic pavement distress detection method system using Retina Net model and U-Net model.This method can automatically detect longitudinal crack,transverse crack,alligator crack,block crack,path block and path strip.And automatically measure length,and area of distress.Specific research content includes: construction data set for pavement distress detection model training and data set for pavement crack contour extraction model training,respectively;An automatic pavement disease detection algorithm based on the Retina Net model is proposed.This model first determines a priori bounding box suitable for pavement disease detection through K-means clustering analysis algorithm,and then uses Res Net and feature pyramid net to extract feature map,Finally,two FCN subnets with the same structure but no shared parameters are used to complete the classification and position regression tasks of the bounding box;An automatic algorithm for pavement crack contour extraction based on U-Net model is proposed.The model includes two parts,the contraction path and the expansion path.The contraction path is used to obtain the feature information of the original image,and the expansion path is used to accurately detect position of the cracks,the model also incorporates the Convolutional Block Attention module attention mechanism module to reassign weights of the feature map to improve the detection accuracy of the model;After that,the width,length,area of the distress were extracted,and grade classification and comprehensive evaluation of pavement damage degree were carried out based on the test results.Results indicate that the m AP of the Retina Net is 62.66%.And the overall precision of U-Net is 90.17%.In conclusion,research in this dissertation is quite effective and practical.The proposed methods has high generalization ability and accuracy,can provide a new idea and method for road surface disease detection,improve the degree of automation of road surface disease detection,and has important practical significance for scientific maintenance decisions of highway management departments.