Research On Crack Detection Method For Pavement And Tunnel Lining Surface Images Based On Deep Learning

In 2022,the Ministry of Transport released the“Highway‘14th Five-Year’Plan”,which clarifies the general idea of China’s road traffic development in the"14th Five-Year Plan"period,not only to increase construction efforts,but also to pay attention to the development of management,maintenance,operation and other aspects.As the national highway system continues to expand and come into service,many of the original roads and tunnels have begun to appear in varying degrees of disease and damage,if these diseases are not repaired in a timely and reasonable manner,it will cause huge hidden dangers to traffic safety.Cracks as one of the main diseases,are the focus of highway and tunnel maintenance.This thesis focuses on the detection of cracks in tunnel lining surfaces and pavement cracks.The complex environment of pavement and tunnel lining surface,the interference of light intensity,shadows,and stains,the small percentage of crack pixels,and the unbalanced distribution of feature weights all contribute to the inability to effectively extract crack features.Tunnel lining surface crack detection more often uses a multi-step digital image processing method of percolation pre-processing,lining seam detection,marking,removal,and denoising to get the final result,which has low detection efficiency,and coupled with complex backgrounds such as lining seams and crack morphology and grayscale features,it is easy to cause detection interference and redundant calculations.In recent years,deep learning has advanced significantly in the field of machine vision,and its recognition accuracy has surpassed that of digital image processing and machine learning algorithms based on artificially designed features.Aiming at the problem of difficult detection of cracks in pavements and tunnels,the design of a fast intelligent crack detection method based on deep learning will provide an important guarantee for the safety assessment and disease prevention of roads and tunnels.The main work of this thesis is studied as follows:1.Aiming at the problems discontinuity and irregular crack morphology cause interference to detail detection and leakage problems on pavement crack detection,two optimization methods based on crack feature extraction are proposed in this thesis.Based on the morphological characteristics of cracks,the Mixed Pooling Module（MPM）is proposed to obtain continuous crack information,optimize the spatial structure information at low levels using Spatial Attention（SA）,capture the context information at high levels using Channel Attention（CA）,improve the efficiency of using crack features,obtain richer feature information,and fuse the information at each level to obtain accurate crack detection images.Based on the symmetric encoder-decoder network structure U-Net is improved by using the residual module instead of the original convolutional block,using cascaded dilate convolution in the middle layer of the network to obtain a richer receptive field,obtaining global information,using Deeply Supervised strategy to improve the learning ability of each side of the network,and fusing the information of each side to obtain the final prediction results.The final ODS values of 0.861 were obtained on the Deep Crak dataset and 0.794 on the CFD dataset.2.Aiming at the problems of discontinuous crack grayscale,inconsistent features,and unbalanced background pixel occupancy,a pavement crack detection method based on enhanced convolution and dynamic feature selection is proposed.The Enhanced Convolution Modula（ECM）adds symmetric horizontal and vertical convolution blocks to the conventional convolution to obtain richer feature information,and uses structural reparameterization to fuse the newly added convolution blocks in the inference stage,which can obtain performance improvement without increasing the inference time.Cracks are irregular in shape and mostly present long strip structure,conventional convolution can only cover a small portion of the crack region,resulting in inefficient feature extraction,so strip convolution is proposed to capture the long distance information of cracks.To address the situation that features at different scales in the network contribute differently to the final detection results,an attention-based Dynamic Feature Fusion（DFF）strategy is used to learn the fusion weights of multi-scale features adaptively.The model finally achieved an ODS value of 0.872 on the Deep Crak dataset,0.788 on the Crack500 dataset,and 0.863 on the CFD dataset.3.Aiming at the problems of tunnel lining surface crack highly similar to lining seams,complex background noise,and low detection efficiency,a tunnel lining surface crack detection method based on Mixed Attention（MA）and multi-scale feature fusion is proposed in this thesis.The method takes advantage of the characteristics of lining seams and cracks in tunnel lining surface images,and designs an efficient mixed attention module that can effectively distinguish lining seams and crack features.The mixed attention module combines channel attention and decoupled spatial attention.The channel attention adopts the Efficient Channel Attention（ECA）module,and the spatial attention is aggregated directly along both horizontal and vertical directions,which can effectively capture the long features of cracks or lining seams.Meanwhile,in order to better refine the crack features,a receptive field enhancement module is proposed to increase the global information of the network,and a new multi-scale feature fusion method is used to directly integrate the rich semantic information extracted from the receptive field enhancement module into the low-level features,avoiding the degradation of semantic information caused by network stacking.The model finally achieved the best F₁ value of0.704 on the proposed Tunnel200 dataset.4.Aiming at the problems of problems of complex crack detection network model,large computation and slow detection speed,a lightweight crack detection algorithm based on efficient convolutional reorganization is proposed.The method proposes a new Split and Exchange Convolution（SEConv）,which splits the feature map into two parts with high and low resolution and exchanges the feature information,while using group convolution and point convolution to reduce the number of model parameters,filter redundant information,and then reorganize.While reducing the complexity of the model,the information exchange of cross-scale features is carried out to avoid the degradation of detection accuracy,and the transformation of feature scales is achieved by using convolution,pooling or up sampling operations according to the scale size of the current features,and then the feature information of different scales is fused.The proposed module can efficiently integrate feature information in stage and cross stage,while filtering out some unimportant information.Experiments show that the method is able to maintain high detection accuracy even at lower model complexity.The model has only1.3M number of parameters and 8G（Floating Point Operations,FLOPs）,while achieving F₁values of 0.867 on the Deep Crack dataset,0.737 on the Crack500 dataset,and 0.811 on the CFD dataset.