Image Detection Of Crack Diseases In Subway Tunnels Based On Deep Learning

Subway tunnel transportation plays an important role in modern urban rail transport system.With the urbanization process continues to advance,a large number of subway tunnels have shifted from the construction period to maintenance period.Because of various internal and external factors such as time,deterioration of material properties,other engineering construction and climate conditions,there occur numerous problems in the tunnel lining structure,among which lining cracks are the most common and the most serious one.It is important to detect these cracks in time and efficiently for the later maintenance work.In order to solve the problems of low efficiency and high cost of traditional manual detection,a method of crack detection in subway tunnel based on deep learning is proposed in this thesis,which applies Deep Convolution Generative Adversarial Networks(short for DCGAN)to realize data expansion,YOLOv5 and digital image processing technology to achieve the intelligent identification of subway tunnel cracks.The main work of this thesis is as follows:(1)Firstly,the image of subway tunnel lining is preprocessed and data enhanced to construct the image data set of subway tunnel crack.(2)As to the shortcoming that traditional data enhancement methods are mostly supervised and need to be set artificially,this thesis uses unsupervised generated countermeasure network to generate tunnel crack images,which is based on DCGAN network model framework,and the self-attention mechanism is introduced to extract the local and global features of the image.At the same time,large-scale tunnel crack image is generated through improving the depth of the network,and the loss function is improved,which uses Wasserstein distance to replace JS divergence,in an effort to enhance the stability of the model.The improved DCGAN is proved to be superior to the original DCGAN in the quality and diversity of tunnel crack images.(3)Compared with the current mainstream target detection algorithm,the YOLOv5 algorithm based on PYTORCH framework is selected for the target detection task of subway tunnel image cracks,and the improvement of YOLOv5 algorithm is proposed.In the K-means algorithm of adaptive anchor frame generation phase,GIOU is used instead of IOU as the evaluation index of similarity between targets,which improves the precision of model detection.The attention mechanism is introduced into Backbone network structure to improve the detection rate of small target cracks.Finally,it is proved that the improved model is 4.9% more accurate than the original model in the average accuracy of the enhanced data set.