Research On Pavement Crack Recognition Method Based On Two-stage Convolutional Neural Network

With the rapid development of highway transportation,it is becoming harder and harder to maintain pavements.Pavement management departments need to quickly and timely master road surface information.The artificial methods cannot satisfy the basic requirements of the development of the road.At the same time,for the large scale road distress images,the methods based on image recognition cannot also meet the application of the large scale data because these algorithms suffer the noises from the road image,the limitation of feature extraction methods,large amount of image data and so no.It has been gained better results from convolutional neural networks in big data processing such as images,voices and so on.However,convolutional neural networks need take a longer time for the large scale data,which seriously influences the further application of convolution neural networks in pavement crack recognition.In this paper,we researched the accelerating training methods of convolutional neural network and applied it in pavement crack recognition.The main work is as follows:Firstly,a two-stage convolutional neural network model was presented based on the principle of human brain repeated memory.In the model,the convolutional neural network was repeatedly trained through randomly selecting some samples in the first phase.Then a network model was obtained.In the second stage,the network model from the first stage was taken as the initial weights of the network.Then the network was optimized with all samples and the final model was achieved.Secondly,we studied the method to determine the number of the training samples in the first stage and the selection of convolutional kernels.To determine the number of the training samples in the first stage,the network was trained on the whole data set and the elbow of the error curve was taken as the number of samples of the first stage.Experimental results on the pavement crack data showed the method was effective.Simultaneously,we compared the effect of different convolutional kernels on the accuracy through experiments and decided the size of the convolutional kernels.Lastly,the two stage convolutional neural network was constructed and applied in pavement crack recognition.At the same time,the network was compared with the traditional recognition methods.Experimental results showed that the recognition accuracy was improved and the train procedure was greatly accelerated through applying the model in pavement crack recognition.