Research On Identification And Analysis Methods Of Highway Tunnel Lining Crack Based On Deep Learning

With the continuous improvement of transportation infrastructure in China,the scale and quantity of tunnel construction are constantly expanding,and the task of tunnel inspection is gradually increasing.Among them,tunnel lining cracks,as one of the early manifestations of tunnel diseases,have an important impact on the safe operation of tunnels.However,the current tunnel lining crack detection technology and equipment in China are relatively backward and cannot meet actual needs.Traditional tunnel lining crack detection and identification mainly rely on manual inspection,with low efficiency and accuracy,serious missed detection,and difficult to meet the growing demand for tunnel inspection due to the increasing number of tunnels.Although digital image processing technology has improved compared to manual inspection,it still has problems such as low detection efficiency,low accuracy,multiple model parameters,and large volume.It also has strong targeting and subjectivity,and its effect is limited when processing disturbed images,making it difficult to have good generalization and robustness,and the detection effect is not ideal.With the rapid development of deep learning technology,using deep convolutional neural networks to quickly and accurately extract target feature information has been widely studied.Therefore,based on the National Key R&D Program sub-project "Intelligent Perception Mechanism and Data Analysis Method for Road Infrastructure",this thesis conducts research on the identification and analysis method of cracks in highway tunnel linings based on deep learning,which has important practical significance.Based on this,this thesis mainly carries out the following related research:(1)In view of the low detection accuracy,high false detection rate and missed detection rate,large volume and multiple parameters,and low detection efficiency of current digital image processing technology for tunnel lining cracks,this thesis,based on the original YOLO v5 architecture,adds a new low-dimensional scale to the detection layer,integrates SE attention mechanism into the feature extraction layer of backbone network and improves the CSP structure.An improved YOLO v5 depth target detection model is proposed for the intelligent detection of cracks in highway tunnel lining.The experimental results show that when using the highway tunnel lining crack detection method YOLO v5-IBX proposed in this thesis to detect cracks,the average precision is98.6%,which is 10.1% higher than the original model,and the detection time is 0.021seconds/frame,with a weight of only 25.2MB.While improving the detection accuracy,the model detection time and network lightweight are guaranteed,and the detection model is optimized.It also has a low false positive rate and missed detection rate,which are 2.2%and 2.3%,respectively,both of which are within the allowable detection error range of less than 10% specified by the standard requirements,and can effectively distinguish between real cracks and false cracks,providing a more effective detection solution for highway tunnel lining crack detection.(2)In response to the problems that the target detection algorithm only detects the existence of cracks and does not go deep into pixel level recognition,as well as the low recognition accuracy,large number of parameters and low recognition efficiency of the classical semantic segmentation model,this thesis introduces the Res Net50 bottleneck residual structure and CA attention mechanism in the coding stage based on the original U-Net network architecture.A 1×1 convolutional optimization network structure is introduced in the decoding stage,and an improved U-Net deep semantic segmentation model is proposed for intelligent recognition of highway tunnel lining cracks.Experimental results show that when using the crack segmentation method U-Net-IBX proposed in this thesis to segment the lining cracks,the average pixel accuracy reaches92%,which is 7.9% higher than the original model,and the training time is 0.34 s,with a weight of only 78.6MB.While improving segmentation accuracy,the network is lightweight.Moreover,the comparison of the extracted crack results of the tunnel lining between the improved method and the original method shows that the cracks extracted by the improved method are more continuous,rich,and accurate,providing a more effective segmentation solution for the feature extraction of cracks in highway tunnel linings.(3)In view of the low efficiency,low accuracy,and unsatisfactory results of mainstream crack quantification methods,this thesis uses the Ostu maximum inter-class variance method,morphology processing,the improved Zhang fast parallel thinning algorithm,and chain code tracking to extract the crack skeleton from the semantic segmentation image of the tunnel lining crack,calculate the geometric parameters such as the length,width,and area of the crack,and compare them with the scale coefficient and actual measurement values.The results show that the mean length extraction error of this thesis’ s algorithm is 4.7%,and the mean width extraction error is 2.8%,both of which meet the maximum allowable detection error in the standard lining crack quantitative assessment criteria,and can effectively achieve automatic calculation of cracks,providing a more effective quantification solution for the geometric feature parameter calculation of highway tunnel lining cracks.