Research On The Detection Method For Road Surface Defects Based On Machine Learning

The road is not only the basis for a country to develop its economy and improve its people’s livelihood but also the embodiment of its comprehensive national strength.With the ever-growing road mileage in China,the road management and maintenance are becoming increasingly significant.Therefore,the detection of road surface defects has become a research hotspot.Although many defect detection methods have emerged at present,there are many matters,such as low recognition accuracy and automatic level,slow detection speed,and lack of universal algorithms.Based on the traditional machine learning method and deep learning method,this paper takes the road surface defects as the research target,and is devoted to improving the automatic level and accuracy of detection.It focus on the feature extraction algorithm of road image,dimension reduction algorithm,defect recognition algorithm and completes the corresponding program design.In addition,a road surface defect detection system has been designed for realizing automatic detection.The detailed research contents are as follows:(1)Based on the traditional machine learning method,the texture features of road surface defects are extracted from three aspects: gradient image,gray image,and Gabor transform image.Besides,Hu moment invariant feature also is extracted form the defect image.Based on these features,the feature set representing road surface defects is established.In order to avoid the interference of redundant information in high-dimensional feature and reduce the complexity of calculation,isometric mapping,laplacian eigenmaps,transfer component analysis are employed to remove redundancy,reduce the dimension and improve the calculation efficiency.Recognition models are constructed by the probabilistic neural network,k-nearest neighbor and random forest algorithm,and used to identify the category of defect.According to the experimental results,the method that employ transfer component analysis to reduce dimension and random forest algorithm to build recognition model can get an over 99% recognition accuracy.(2)Based on the deep learning method,the convolution neural network widely used in the image field is selected to extract the defect image features and identify the category of defect.Convolutional neural network has a good advantage of feature self-extraction,which avoids the interference form artificial extraction and omits the artificial steps.Three pre-trained convolutional neural networks,squeezenet,resnet101 and densenet201,are selected and make them adapt to the task of road surface defect recognition through corresponding modification.According to the experimental results,compared with squeezenet and resnet101,densenet201 network is better adapted to the task of road surface defect identification in this paper,and its recognition accuracy is more than 99%.(3)On the basis of good practical results,combined with the defect identification methods in the traditional machine learning((1))and the deep learning((2)),a road surface defect detection system is designed based on Matlab GUI and SQL server.The system possesses these functions of automatic defect identification,the display and storage of image information,image query and retrieval.It can realize accurate automatic recognition of road surface defects and achieve the purpose of automatic detection.