The Study On Classification Number Under Deep Learning And The Application Of Classification Of The Cavity Beneath Pavement Slabs Based On CNN

Cement concrete,as the major component of highway pavement,has strong and durable characteristics.But with the influence of some natural factors,the cement concrete pavement will also be damaged to different extents.The cavity beneath cement concrete pavement slabs is the most common one.However,the current classification of the cavity beneath cement concrete pavement is determined by experience lack of objective researches on automatic classification.Firstly,this paper proposes a method to determine the number of classifications based on Akaike Information Theory Criteria(AIC criterion).The main contents are as follows: the selection of local Gaussian kernel function in the similar matrix,the construction of normalized Laplacian matrix,and the updating of AIC criterion and MDL criterion according to the normalized Laplacian matrix to make it suitable for the calculation of the number of data sets.Secondly,this paper uses GA index to evaluate the classification accuracy of the detection dataset of the cavity beneath cement concrete pavement under different classification numbers,and to judge the classification number.Finally,this paper also adjusts the mechanism of VGG-16 network: removing two fully connected layers to reduce most of the parameters in the network structure adjusting the convolution kernels neurons of the convolution unit 5 to700 per neuron which greatly reduce the amount of data to be saved;this paper adopts data amplification technology to increase the labeled sample data according to the characteristics of detection data of the cavity beneath cement concrete pavement and adopting transfer learning method for the pre-training of network model aiming to significantly improve the recognition accuracy of the VGG-16 network model.Through evaluation of GA index,GA index of the three categories was 0.9367 while GA index of the four categories was 0.9883.This paper uses VGG-16 network to train and test the data set.In addition to the above-mentioned two classifications,it also introduces 5 categories for comparison.The results show that the recognition accuracy of VGG-16 network of the three categories is 93.82%;the recognition accuracy of VGG-16 network of the four categories is 97.54%;the recognition accuracy of VGG-16 network of the five categories is 83.61%.