Study On Detection Of Pavement Void Disease Based On Deep Learning And Ground Penetrating Radar

Road void disease is a kind of harmful structure disease,which directly affects the life and performance of roads.When serious,it may cause depression on the surface of pavement and cause serious loss of life and property.Because of the strong concealment of road void disease,traditional detection technology alone cannot meet the requirements of modern road construction.At present,in the process of detecting road void disease in China,there are some problems,such as non-destructive detection,low accuracy,low automation,and so on.In this dissertation,void detection based on deep learning and ground penetrating radar technology is studied.The main contents of this dissertation are as follows:(1)Study on GPR data characteristics of road void diseases.The identification and detection of road void disease are calculated based on the ground penetrating radar(GPR)collected point matrix information and the GPR profile picture generated by data processing.In this dissertation,through automatic gain compensation,void slicing,pixel gray value mapping and Gamma data enhancement,the point array data collected by radar is transformed into image data suitable for deep learning network learning.(2)Road void disease identification based on multi-scale fusion and Attention Mechanism.For the problems of low accuracy and small target in ground penetrating radar(GPR)converted image,a multi-scale fusion network structure is proposed to make full use of the context feature map information to improve the representation ability of the feature map.A new attention mechanism is also designed to convolute and compress global pixels while guaranteeing computational load,to weight the original feature pixels,reduce noise interference,and improve the network’s ability to grasp the global key pixel information.Combining the above two innovative designed convolution neural network structures,the training and testing of road structure void disease dataset recognition can be significantly improved compared with other mainstream networks.Finally,the experimental results on the public datasets cifar10 and cifar100 have also achieved high accuracy,which further proves the universality and practicability of the attention mechanism.(3)Locating road void disease based on Faster R-CNN(Faster Region Convolutional Neural Network).To solve the problem that the convolution neural network is not accurate and can not locate the void location,this dissertation designs Faster R-CNN with different combinations of structures for structure optimization training,and realizes the detection and location of road void disease based on GPR profiles.Finally,the detection method of road void disease based on Faster R-CNN is compared with other mainstream detection methods,and the results show that this method has certain advantages.(4)Automatic road void disease detection system based on Web.This dissertation participates in the design of an automated road void disease detection system based on Web.For the problem of slightly higher miss-report rate in void recognition detection,this dissertation uses the joint elimination algorithm to test,further improves the accuracy and reduces the miss-report rate on the basis of the original recognition.Finally,the system operation effect diagram is given to verify the feasibility of this automated void detection.