Research On Pavement Crack Detection Algorithm Based On Deep Learning

Road maintenance is an indispensable part of the daily maintenance of public transport system,and pavement detection is an important step of road maintenance.As the most common pavement damage,pavement cracks have become the most important detection object in many road detection tasks.The detection of pavement cracks in different environments can promote the road administration departments and drivers to understand the real situation of the road in a timely manner,provide scientific and effective data support for the formulation of subsequent road maintenance plans,and is of great significance for the improvement of road traffic safety index.The deep learning technology learns the sample features through complex nonlinear calculation,and can obtain better detection results when there is a large amount of data.Therefore,based on depth learning technology,this paper proposes a trainable pavement crack detection algorithm to accurately locate pavement cracks with complex background lines.First of all,the thesis studies a variety of commonly used image processing and related depth learning technologies,makes a detailed theoretical analysis of convolutional neural network and residual network,and gives the optimization idea of pavement crack target detection.Secondly,based on the digital image processing method,the road image is grayscale transformed,and the road markings in the road image are removed by using the gray-scale threshold segmentation method.The road image is denoised by means of mean filtering,Gaussian filtering,median filtering and adaptive median filtering,and the processing results of different filtering methods are compared.Through the evaluation of Peak Signal-to-Noise Ratio(PSNR)index,the PSNR value of adaptive median filtering is 30.3,which is the best compared with the other three filtering methods.In this paper,the image enhancement method is used to re process the pavement image,and the pavement image processing results of Histogram Equalization(HE)algorithm and Clip Limit Adaptive Histogram Equalization(CLAHE)algorithm are compared and analyzed.It is pointed out that CLAHE algorithm improves the contrast between the crack target and the background noise,and has better effect on the pavement image enhancement processing.Thirdly,a new pavement crack detection method based on Multi-Area Segmentation Combine(MASC)is proposed.CRes-Net+FPN network is used to extract image features,and the convolution in the basic network is replaced by deformable convolution to solve the problem of incomplete feature representation caused by unknown crack shape,A new MASC pavement crack detection model is proposed in this thesis.The pavement crack information in the image is more comprehensively represented by the whole,core and border multi regions.The improved residual network multi region representation method can effectively solve the problem of false detection caused by non crack texture,The border area information is used to monitor the prediction of the core area to remove the cross noise generated by the edge of the core area and the crack texture border information,making the prediction of the crack core area more accurate.Finally,the thesis uses the Crack Forest dataset to conduct ablation experiments on the MASC pavement crack model,determines the loss weight through testing,and uses PANet,CE Net,U-Net,HED,RCF,ACNet and the MASC algorithm proposed in this paper to detect pavement cracks respectively.The test results show that the three indicators of the MASC algorithm in this paper have good performance,the recall rate has increased by about 1%,and the F value has increased by 1.1%compared with the next best RCF algorithm.In addition,the validity of the pavement crack detection algorithm in this paper for complex background detection is verified through the preliminary data of Street dataset test.Compared with the methods CE Net,PANet,ACNet and Mask R-CNN,the accuracy rate is improved by 6.6%,and the final F value is improved by 3.2%.The detection accuracy rate tends to be stable when the number of iterations exceeds 4000 according to the comparison of detection accuracy rate.Under the same iteration number,the detection accuracy rate of MASC algorithm is about 4% higher than that of other algorithms.The experimental results show that the MASC method proposed in this thesis is more accurate and effective.