| Highway construction is the focus of China’s infrastructure construction.As the urbanization process is accelerated,it also brings huge traffic pressure.The quality of cement road surface is also highlighted,because of the erosion of rain,the role of vehicle load,and the difference of construction methods will affect the cement road surface so that the cement road will produce the corresponding disease.Therefore,regular detection of cement road surface will be the focus of road disease control work.Although GPR technology has been widely used,GPR image recognition technology still has shortcomings.Based on the detection of cement road disease by groundpenetrating radar,this paper uses GPRMAX2 D software to simulate the characteristics of cement road disease pictures.Yolo v3 is used to automatically recognize the disease image and improve the detection accuracy.The method of automatic recognition of cement road damage image is studied.This paper analyzes the causes and characteristics of cement road diseases.Understand the pavement structure,the spatial location of roadbed disease,geometric shape,formation reasons,and different parameters of different diseases.GPRMAX software is used to carry out numerical simulations of different cement road diseases.The images collected by GPR contain noise due to the influence of various factors.The median filter,mean filter,and bilateral filter are used to de-noise the echo collected by GPR,mainly to filter out salt and pepper noise.In this way,clutter and noise can be effectively suppressed,thus improving the quality of the collected disease images.This topic through the analysis of neural network technology,the use of its selflearning,to find the characteristics and rule of actual samples,can according to the characteristics of the training acquisition,predict the new sample,the classification of parallel computing can be independent,improve operation training speed,etc advantages.YOLO was used as the basic framework to train the data set.The convolution mechanism of each module of YOLO v3 is described in detail,and the principle of v3 network recognition is introduced.To improve the recognition accuracy,the recognition loss values under different prediction boundary boxes and different learning rates were analyzed.The collected GPR echo samples were de-noised using mean filtering,median filtering,and bilateral filtering clutter suppression methods,and then put into the trained YOLO v3.By comparing the accuracy of v2 and v3 in the recognition of various diseases,the effectiveness and superiority of v3 in the identification of disease areas were obtained. |
