Research On Pavement Crack Detection Method Based On Convolutional Neural Network

After China has experienced large-scale highway construction,the speed of highway construction is gradually slowing down,and a huge demand for pavement maintenance is ensuing.Pavement cracks are one of the common manifestations of pavement damage,and effective detection of them is undoubtedly of great significance to pavement maintenance.In order to solve the problems of low accuracy and poor universality of traditional pavement detection,this paper takes pavement cracks as the research content and uses convolutional neural network technology to design a new network model for crack classification detection.The specific work is as follows:(1)Marking and classifying the collected crack images,the sample is specifically divided into horizontal cracks,vertical cracks,turtle-shaped,block-shaped,repaired and crack-free pictures.The sample images are divided into training set,verification set and test set in a certain proportion,and the road crack image database required for the experiment is established.(2)Designed a convolutional neural network architecture suitable for crack classification and detection.Firstly,the basic network model is designed,the original Alex Net network and Google Ne network are thoroughly explored,and the characteristics of the extracted network model are analyzed.Based on this theory,the designed model is improved,and the idea of Inception module is introduced to change the topology of the original network.Convolution size,Dropout layer is added,a new crack classification network model is proposed.(3)The optimization strategy of the network model is set,and the corresponding optimization strategies are proposed for various hyperparameters of the convolutional neural network,including learning rate setting,batch amount selection,regularization parameter setting,and optimizer selection.(4)Based on the Keras deep learning framework,the designed basic network model and the improved network model are trained,and the comparison and verification show that the latter network model has improved the accuracy rate of the basic network model by 9.5% under the same experimental conditions,and then According to the proposed model optimization strategy,the improved network model is optimized.After a large number of comparative experiments,the crack classification network model with high accuracy is finally trained.The highest accuracy rate can reach 91.6%.