Research On Concrete Crack Detection Based On Deep Learning

In recent years,with the development of infrastructure,concrete as its important construction material plays a decisive role in the national economy.However,due to various factors such as temperature changes,cracks will occur during the use of concrete structures,and the appearance of cracks will have an impact on the strength and durability of concrete structures.So the crack detection of concrete becomes an important work.Traditional concrete crack detection methods mainly rely on artificial visual detection and machine vision detection,but these methods are inefficient and subject to subjective factors.The concrete crack detection method based on deep learning can realize automatic crack detection by training the deep neural network model.This method is helpful to discover cracks in concrete structures in time,ensure the safety and service life of concrete structures,and has important practical application value.Aiming at the research on concrete crack detection,this thesis starts from two aspects of crack identification,classification and detection and positioning,and respectively carries out optimization research on concrete crack identification,classification and detection and positioning.Specific research contents are as follows:(1)Establish a concrete crack dataset,enrich samples through two channels:online open source acquisition and manual collection,and perform cropping and normalization operations on the acquired images.In order to enhance the number of samples,data enhancement methods are used to expand.On the issue of crack detection,the Labellmg software is used to label the cracks in the dataset,and then the concrete crack identification classification and detection location dataset are established respectively.(2)Aiming at the problem of concrete crack identification and classification,a concrete crack classification model based on improved AlexNet is proposed,AlexNet_SE_SVM.The optimization starts from AlexNet network structure,including adjusting convolutional layer structure,introducing batch normalization and modifying activation function.The SENet attention mechanism is added to enhance the feature extraction of fine cracks.The feature information extracted from the network is used as the input of SVM classifier to realize the crack classification.Experiments show that the optimized AlexNet_SE_SVM model has good classification performance and improves the accuracy of crack classification.(3)Aiming at the problem of concrete crack detection and location,a concrete crack detection model based on improved YOLOv5s is proposed.In order to better match concrete image data set,K-means++algorithm is used to cluster prior boxes.The backbone feature network CSPDarknet53 of YOLOv5s is replaced with ShuffleNetV2 to achieve network lightweight.In order to better detect fine cracks in complex backgrounds,the CBAM attention mechanism is added to the backbone network to better focus on features at different locations.The FPN+PAN network structure in Neck module is improved to BiFPN network to further strengthen the feature fusion.Experiments show that the optimized YOLOv5s_K_SCB model has higher detection performance and speed.