Research On Core Algorithms For Intelligent Analysis And Identification Of Road Surface Defects

As one of China transportation infrastructure projects,road has a great impact on citizens' lives and national economy.Road surface,the important factor affecting traffic development,has become the main object of road maintenance.Therefore,timely and efficient detection of road surface defects is a key to ensure the quality of road.In order to analyze and identify the road surface defects intelligently,this paper undertakes the research on related algorithm based on machine vision technology.According to the principles of feature extraction,feature selection,traditional recognition model and deep learning,the corresponding programs are designed and the models are built.After the recognition of roads surface defects images,the classification performances of all models are compared by the relevant classifier indexes.The main research contents are as follows:(1)For extracting and expressing the feature information of pavement defect images as completely as possible,this paper makes different representation of pavement defect image from different angles.Starting from the gray statistical image,the gray histogram algorithm,the gray difference algorithm and the gray gradient algorithm are used to extract the features separately;starting from the statistical analysis of change rules of the gray matrix,the gray co-occurrence matrix algorithm,Gabor algorithm and Tamura algorithm are used to extract the features separately;starting from the view of image morphology,boundary geometric feature extraction algorithm,boundary Fourier description algorithm and moment invariant description algorithm are used to extract features respectively.(2)Saliency analysis of extracted features is necessary so as to prevent multi-feature redundancy leading to low discrimination.After statistical analysis of the variance distribution of the features,Kruskal-Wallis algorithm in non-parametric test is selected to verify the significance of the features.According to the verification results,taking advantage of LDA algorithm,MI algorithm and PSO algorithm complete the selection of effective features to simplify the representation and accelerate the recognition operation.(3)Classifier is treated as the crux of road surface defect images recognition.Based on the theories of BP neural network,support vector machine and extreme learning machine,this paper constructs the corresponding classification models.After contrast experiments,the best traditional recognition model is obtained by comparing the established classification performance index of each model.(4)Research on deep learning models with the ability of processing original images instead of feature extraction and feature selection which include deep belief network,stack sparse self-encoder,and VGG convolution neural network are developed when the corresponding programs are designed according to the principles.After prediction,the identification model with the best identification performance is obtained through comparing the recognition accuracy and other performance indicators of each model.