The Study On The Recognition Of The Hidden Defects Of Expressway Pavement Based On The Convolutional Neural Network

Hidden defects of highway is an important factor leading to large-scale damage of highway,which is directly related to the service life of highway.The general detection methods is hard to locate road defects accurately,because of the strong concealment and random distribution.Therefore,a accurate and efficient detection method of hidden defects on the highway has become an important basis for road maintenance.It is urgent to develop a set of automatic or semi-automatic hidden defects identification system,because of many limitations of current detection and recognition technology,such as low efficiency,artificial dependence.According to practical situation,the author provides a new system based on GPR and Deep Learning to identify hidden road defects.The main researches is as follows:(1)Based on the latest theory of highway detection at home and abroad,the author compares and analyzes many non-destructive detection technologies,such as ultrasonic detection,CT detection,optical fiber detection and GPR detection,and decides to choose GPR with strong anti-interference ability,high detection accuracy and convenient operation as the detection equipment for the hidden defects data collection of highway in this paper.(2)According to the propagation law and detection principle of GPR electromagnetic wave in the medium,the author analyzes various factors affecting the detection accuracy of GPR and puts forward corresponding solutions.At the same time,the author explains the characteristics of radar response map of three kinds of highway hidden defects of crack,void and subsidence,and explains the GPR parameters that need to be adjusted when collecting the hidden defects sample data.(3)The collected sample images is preprocessed firstly,and then the Gauss filtering is used to filter the image,and the image enhancement technology is used to enhance the disease characteristics of the image,according to the noise characteristicsof GPR.Compared with the existing image classification algorithms,the author chooses convolutional neural network(CNN),which has high classification accuracy and strong image processing ability,as the basic algorithm of image processing in this paper.(4)The data set is divided into training set and test set according to a certain proportion.By analyzing the principle of CNN and deepening the depth of CNN to extract higher-level abstract features,the classification model of CNN is established.By studying the inception module,the author introduces the depth separable convolution to solve the problem of too much parameter and calculation in CNN due to too deep network,and establishes the depth separable CNN model.At the same time,in order to prevent the distortion of input image feature extraction caused by the reduction of parameters,the standard convolution is selected for the shallow layer of network.(5)The author classifies the defect images by using model.In feature extraction,support vector machine will miss some useful features with a classification accuracy of 88.333%;traditional neural network can not extract advanced features with a classification accuracy of 89.6% due to the requirements of parameter quantity;convolution neural network has a deep network with comprehensive features,with a classification accuracy of 92%,but training time of 10 hours with a large number of parameters.However,the deep CNN optimizes the network structure,the classification accuracy is 94.62%,and the training time is 5 hours,which not only reduces the training period,but also improves the classification accuracy.