The Research And Implementation Of Defect Identification Of Highway Pavement Based On CNN

In view of the single crack texture feature used in traditional pavement defect detection research,it is impossible to effectively identify other pavement defects under complex road conditions.On the one hand,potholes are common types of road diseases that exist in road damage,and need effective methods to identify and analyze them.On the other hand,in addition to intact road sections and defective sections,road waste will appear.The traditional road defect method is prone to misjudgment,due to disturbances caused by objects,sunlight,and sidecars.The identification efficiency is low and the effect is not satisfactory.In the past,the method of judging the road surface condition by manual work is not suitable for the requirements of high-quality road maintenance tasks,and the road surface defect identification method requires more efficient detection means.The deep learning method has been widely studied and applied in the field of image and machine vision.It is inspired by the deep learning method.This paper uses the deep convolutional neural network as the research method,and studies the convolutional neural network CNN with various pavement defects as the object.The application of pavement disease defect identification,using a well-trained deep convolutional neural network,completes the automatic extraction of road pavement defect features,and realizes the classification of cracks,pits,road debris and intact pavement.Based on the Caffe design framework,the original VGG model is improved and the parameters are optimized.The dataset enhancement method is used to expand the original data,and the data is used for the training of convolutional neural networks.The test results show the feasibility of the research method.The research work of this paper is mainly reflected in the following parts:(1)The characteristics and application of existing road defect detection methods at home and abroad are fully analyzed.The method of classification of road surface defect data by deep convolutional neural network is studied for complex road surface defect detection requirements.(2)A large number of datasets of pavement defect types were collated and manually labeled.In nearly 10,000 pictures,cracks,pits,road debris and intact road images were screened.For the sparse data samples,data set enhancement technology was used to expand the original data samples.Image preprocessing methods using global contrast normalization and data set enhancement techniques enrich the defect features and improve the generalization ability of neural networks.(3)A variety of deep convolutional neural network model structures were studied,and VGG was selected as the final research object,and its network structure was optimized to make it suitable for the classification and identification of a variety of pavement defects.The regularization loss and adaptive learning rate are used to reduce the generalization error of the model.(4)The improved neural network model structure was built based on Caffe framework,and the relevant experimental environment was set up.The experimental pavement defect dataset was converted and input into the model for training,and the comparison experiment was conducted with AlexNet model and original VGG model.(5)Based on Visual Studio 2013+MFC,the pavement defect recognition software is designed,which is divided into online and offline detection modes.The software visually displays the recognition effect of the proposed algorithm on pavement defects.