Research And Implementation Of Pavement Crack Classification And Recognition Based On Deep Learning

In recent years,China's highway transportation industry has developed rapidly.By the end of 2017,the total mileage of China's expressways has reached 4,773,500 kilometers,of which the highway mileage is 136,500 kilometers.At present,the maintenance workload of the road surface is becoming increasingly heavy,and obtaining road surface damage information quickly has become the primary task of road maintenance.Crack target is one of the main forms of pavement damage.The traditional manual detection method displays poor universality,low efficiency,and low recognition accuracy when processing massive pavement damage images.Therefore,designing a set of real-time,effective,robust,and high-accuracy detection algorithms has become the core work of pavement crack recognition.In view of the shortcomings of the traditional road surface crack recognition algorithm,this paper analyzes the effects of convolutional neural network(CNN)in large data such as image processing,speech recognition,and semantic analysis,and successfully applies CNN to pavement crack recognition.The main research work of this article is as follows:(1)Data preprocessing for original pavement crack imageFirst,a brightness elevation model is used to homogenize the road surface image to eliminate gray unevenness in the image.At the same time,the noise in the road surface image is effectively reduced,and the contrast between the background and the crack in the image is effectively improved.The road surface image is divided into several sub-blocks of size6464?.The sub-block image is marked and the independent road surface images are divided into a training set,a verification set,and a test set.(2)A convolutional neural network architecture suitable for pavement crack recognition is designedThis article has improved the AlexNet network from CNN's convolutional layer,network topology structure,filter settings,and so on.This article has made AlexNet network to CNN from the network's convolutional layer,network topology,filter settings,etc.To improve,a new convolutional neural network architecture was proposed.Compared with the original AlexNet architecture,this paper has fewer network layers and smaller parameter calculations,which can effectively shorten the sample training time and improve network performance.(3)Developed a variety of CNN optimization strategiesThe process of CNN pavement crack detection is designed.The adjustment strategy of CNN is proposed according to the order of data preprocessing,network model parameter initialization,learning rate setting in training process,activation function selection,and regularization constraints.The preprocessing module includes steps such as image denoising,enhancement,and segmentation.For the particularity of image features of pavement cracks,this paper optimizes the selection of filter size,batch setting,and the selection of learning rate.(4)Training pavement crack classifierIn the process of sample training,this paper presents a data set format conversion for the pavement crack image,and initializes the network with the parameter settings in the Solver.prototxt file.In the end,rapid crack detection on the road surface was achieved.Through a large number of experiments,the CNN network proposed in this paper can achieve efficient detection of crack targets in pavement images.Under the same conditions,the CNN model designed by this paper is less time-consuming than the AlexNet network,and the prediction accuracy rate is higher.The highest accuracy can be up to 96.6%,and it has good engineering application value.