Research On Pavement Distress Inspection Based On Deep Learning And Ground Penetrating Radar

Pavement distresses is one of factors that influence performace and lifetime of highways,especially pavement structure distresses,which show bad effects on highway lifetime,such as reflection cracks,interlayer empty,and uneven settlememt.However,there are several disadvantages of pavement structure detection in China now,such as narrow detection range,low precision,low automatic processes and so on,because the location of the crack is,by definition,difficult to find.It is necessary to do some research on pavement structure detection.Therefore,an attempt to employ deep learning and ground penetrating radar(GPR)technology for asphalt pavement structure distress detection is presented in this manuscript.The main works are listed as below.(1)The research on response waves of pavement distresses in GPR images.The response waves in GPR images of different pavement distresses are the basic of recognition,location,and meansurement.Theoretical calculation,laboratory test,and field test were done to define the characteristics of response waves of reflection cracks,interlayer empty and uneven settlememt.(2)Recognition of pavement structure distress based on convolutional neural network(CNN).A structure design for a CNN for recognizing pavement structure distresses is presented in Section 3.After training and testing,the capacity of recognizing pavement structure distresses was realized.Then the truth grounds were compared with the prediceted results given by the CNN to analyze its precision under different conditions,such as pavement structures and transmitting frequencies.At last,the CNN-based method was compared with other state-of-the-art methods to show it can recognize different distresses in realistic situations.(3)Location of pavement structure distress based on Faster Region convolutional neural network(Faster R-CNN).We designed 30 Faster R-CNNs with different structures.After training and structure optimization,the capacity of automatically locating pavement structure distresses was realized.Then the truth grounds were compared with the prediceted results given by the optimized Faster R-CNN to analyze its precision under different conditions,such as pavement structures and transmitting frequencies.The CNN-based method was compared with other state-of-the-art methods to show it can locate different distresses in realistic situations.At last,the real-time and continuous GPR detection was presented to precision and real-time capability of the optimized Faster R-CNN.(4)Measurement of pavement structure distresses based on Regression convolutional neural network(Reg-CNN).Two Reg-CNNs with different strucutures were designed to realize the automatic measurement of pavement structure distresses.The stability of two Reg-CNNs was vertified by testing under different conditions,such as pavement structures and transmitting frequencies.At last,the Reg-CNN-based method was compared with other state-of-the-art methods to show it can measure different distresses to show its superiority.(5)Reconstructure 3D crack models by the cascade convolutional neural network.A 3D reconstructure method based on cascade convolutional neural network was presented in Section 6.A structure design for a cascade convolutional neural network for rebuilding pavement cracks was presented.After training and testing,the capacity of rebuilding pavement cracks was realized.Then the truth grounds were compared with the prediceted results given by the cascade convolutional neural network to analyze its precision under different conditions,such as pavement structures and transmitting frequencies.At last,the effect of fresnel zone diameter on the precision of 3D models was discussed.