Research On Highway Tunnel Lining High Noise Crack Image Fine Identification Algorithm

With the increasing number of highway tunnels in China,new challenges are posed to the safety inspection of road tunnels.Cracks are one of the most common and important diseases in highway tunnels.Although the technology of automatic acquisition of apparent images of highway tunnels has been relatively mature,the amount of data collected from apparent images of tunnels is very huge,30,000 images need to be collected per kilometer of highway tunnels,and automatic recognition technology of tunnel cracks based on deep network technology needs to be developed to deal with the massive demand of tunnel image detection.Due to the rough concrete lining surface of the tunnel and the apparent pollution caused by the tunnel operation environment,as well as the limitations of the acquisition equipment itself,the acquired highway tunnel lining apparent images have high noise,low image contrast and uneven grayscale distribution,which cause great difficulties for crack identification and extraction.In this paper,for the high noise crack images of highway tunnel lining,we improve the data enhancement algorithm of highway tunnel apparent image from image enhancement algorithm,target detection algorithm,crack segmentation network and crack parameter extraction,and innovatively propose the crack segmentation and key point extraction network of highway tunnel,and realize the fine recognition and parameter extraction of high noise crack images of highway tunnel lining by discussing the actual scale of cracks calculation.The main work and conclusions of the paper are:1.For the problem of low quality of highway tunnel lining images,a new image enhancement algorithm applicable to highway tunnel lining images is proposed based on grayscale histogram equalization algorithm.In order to strengthen the accuracy of the road tunnel crack recognition network,the image quality enhancement algorithm proposed in this paper is used to enhance the training image data,and the effect of this enhancement algorithm is verified by three different crack detection networks,and the experimental results show that the proposed image enhancement algorithm makes the distribution of the histogram more balanced by equalizing the histogram method.It can effectively improve the accuracy of highway tunnel crack recognition,and the crack recognition model based on YOLOv5 has the best evaluation index,which can well accomplish the task of highway tunnel crack detection.2.Aiming at the difficult problem of fine-grained recognition of cracks in road tunnel apparent images,the DeepLabV3 semantic segmentation network is used to study the most suitable deep network structure for the characteristics of road tunnel cracks by modifying different feature extraction backbone networks and different loss functions.The backbone network experiments and loss function experiments illustrate that the ResNet101 network has certain advantages in the highway tunnel feature extraction scenario,while different loss functions have more effects on network training than the backbone network,and the accuracy of crack segmentation can be effectively improved by optimizing the use of Focal loss and mean squared error loss.On this basis,an innovative crack key point extraction module is proposed to realize the extraction of crack key points to assist the fine detection of cracks,and the accuracy of crack identification reaches 90.47%.3.The image pixels,camera accuracy and actual width are discussed by designing image line segment width validation experiments for highway tunnel crack feature extraction needs.The experimental results yield the existence of linear relationship between the number of pixels of cracks in the image and the actual width of cracks under the scenario of low resolution.By studying the skeleton extraction algorithm to realize the skeleton extraction of the crack mask data,calculating the crack directions based on the extracted crack skeleton,realizing the calculation of crack width and length at pixel scale,correcting the actual width of the crack by the pixel-width relationship,and realizing the accurate extraction of crack feature parameters.