Tunnel Disease Detection And Classification Based On Deep Learning

Manual inspection and maintenance has always been the main way to ensure the safety of underground tunnels.With the development and construction of subway tunnel blowout,this method has been unable to adapt to the development trend of intelligent operation and maintenance in the subway industry.In recent years,due to the rise of convolutional neural networks,deep learning-based methods have been favored by more and more researchers in the field of computer vision.Based on deep learning,this paper studies the disease detection algorithm in tunnel scene.In recent years,how to improve the detection accuracy and efficiency of the disease detection algorithm in tunnel scene has become the most concerned problem in the industry.Because the detection algorithm needs to meet the challenge of both classification and location,its classification branch recognition performance is still unable to cope with fine-grained classification tasks.When dealing with cracks and other diseases,relatively similar scratches usually cause a large number of false alarms,which seriously reduces the detection accuracy of cracks.Based on these problems,the main research contents and innovation points of this paper are as follows:(1)An improved RepPoints network architecture based on multi-level cascade is proposed.At present,conventional tunnel disease detection model mostly adopts detection algorithm based on anchor,which brings some design barriers to the improvement of detection accuracy.This chapter first compares the detection performance of anchor detection algorithm and RepPoints algorithm based on point set to represent targets in tunnel disease data set,and verifies the adverse impact of rectangular box representation on some disease detection and some advantages of RepPoints detection algorithm.Secondly,in view of the problem that multiple prediction boxes detect the same target in RepPoints network during disease detection,an improved RepPoints network structure based on multi-level cascade is designed to effectively improve the target positioning effect of sample locations near the center of the disease.Finally,based on the improvement of RepPoints,the performance of disease detection algorithm is further improved by adopting DCNv2(deformable convolution networks v2)and GIOU Loss(Generalized Intersection over Union Loss).(2)A new BET-ResNet(Bottom Enhance Top Residual Network)Network structure was proposed for further screening of detected crack samples to make up for the insufficient detection accuracy of crack targets in the detection network.To solve the problem that the traditional classification network is difficult to obtain the fine granularity classification features with good discrimination between cracks and scratches,BET-ResNet adopts the hierarchical fusion of the bottom features to add details to the top features.At the same time,in order to solve the background interference caused by the increase of the receptive field as the network layers get deeper.Before fusion,channel attention mechanism is introduced for each hierarchical image feature to guide the network to pay attention to the valuable feature information.The recognition performance of BET-ResNet network structure is improved obviously.