Research On Crack Detection Of Concrete Bridge Based On Deep Learning

China is a highly advanced country in bridge construction,and the total number of bridges ranks among the top in the world.Cracks are one of the most important diseases in bridge engineering.Whether they can be detected and maintained in time is directly related to the safety and durability of the bridge.In view of the current bridge crack detection mainly relies on manual,poor safety,strong subjectivity and other shortcomings,the traditional crack detection algorithm accuracy is low,poor generalization ability and other problems.In this paper,a series of researches and improvements are carried out on the bridge crack target detection and contour extraction by using the deep learning correlation algorithm.The main research contents and innovation points of this paper are as follows:(1)Bridge crack image preprocessing.In this paper,the crack of concrete bridge is taken as the research object.The experiment involves the data samples of the crack of bridge with different texture and orientation,which increases the diversity of the training data.According to the specific experimental requirements,in the data preprocessing stage,image enhancement and sliding window technology are used to expand the number of training data.For supervised learning tasks,Labelimg and Labelme image annotation tools are used to manually label crack images and form labels,so as to provide data sets for training deep learning models.(2)Bridge crack target detection based on YOLOV3 model.In this paper,the YOLOV3 model is applied to the automatic detection of bridge cracks.Among them,the original Kmeans clustering algorithm is improved: IOU is used instead of Euclidean distance as the evaluation index of clustering results to make the network model more sensitive to crack targets;Secondly,in the process of model training,we innovatively propose to use Early-Stopping strategy to monitor the changes of verification set loss in order to balance the relationship between training time and generalization error.Finally,the experimental results are analyzed from two aspects: visualization of results and quantification of evaluation index.The experimental results show that the accuracy of the model is 90.26% on the test set,and it has a good crack identification effect.(3)Detection and extraction of bridge crack contour based on image segmentation algorithm.In this paper,a new model of bridge crack contour extraction network is designed based on image semantic segmentation network.The model consists of a down-sampling part,an up-sampling part and a Sigmoid activation function.Residuals are added to the sub-sampling and up-sampling parts to accelerate the model convergence and prevent the gradient from disappearing.Experimental results show that the proposed algorithm is superior to many existing image segmentation algorithms,and has better generalization ability and contour extraction effect.(4)Algorithm acceleration strategy.In this paper,CUDA is used to train and test the GPU acceleration and lifting model.The analysis of experimental results shows that the crack detection time is shorter after using CUDA acceleration,which greatly improves the detection efficiency of bridge cracks.