Study On Detection Algorithm For Road Surface Disease Based On Faster R-CNN

With the rapid development in road network in China,the road safety issues have attracted increasing public concern,therefore,in order to protect people’s life and reduce the loss of national property,the highway infrastructure needs to be maintained effectively and needs prompt handling in case of damage.The occurrence rate of the road infrastructure faults is high,due to the influence of the weather,rolling of vehicles and problems of road quality.For the purpose of dealing with the road infrastructure faults better and faster,the faults need to be founded by the staff in time,but the manual detection cannot find faults in a timely manner.Therefore,there is a hot research issue about how to detect road infrastructure faults accurately and quickly,and the road surface disease detection is one of the hot research issues in road infrastructure faults detection.Automatic detection of road surface disease is a challenging task,the detection of road surface disease based on digital image processing method has achieved a good detection result and is widely used currently.However,the traditional detection method is difficult to attain the desired testing result due to the less effective of image deep level characteristic extraction,with the development of machine learning and deep learning,the detection effect of road surface disease has been greatly improved.Aiming at the problems of low detection efficiency and poor positioning accuracy of prevIoUs detection algorithms,this paper has analyzed the types of road surface disease and image characteristics,and propose a detection algorithm based on deep learning.The main research work of this paper is as follows:Firstly,the research status of road surface disease detection field at domestic and abroad and the research status of target detection field are analyzed,the road surface disease detection algorithms and target detection algorithms at domestic and abroad are compared,the traditional target detection algorithms and the target detection algorithms based on deep learning are deeply studied.On this basis,a road surface disease detection algorithm based on Faster R-CNN(Faster Region-Based Convolutional Neural Networks)is proposed.Secondly,the network structure and working principle of Faster R-CNN are studied.Aiming at the problem of mutual suppression between diseases in the same image caused by the NMS(Non-Maximum Suppression)algorithm in the original Faster R-CNN structure,an improved algorithm of NMS,Soft-NMS is proposed to replace it,thus the redundancy of the detection frame and the rate of missed detection of disease are reduced,and the detection accuracy is improved.At last,in order to further improve the performance of the Faster R-CNN in road surface disease detection,this paper applies data augmentation method and transfer learning method.Data augmentation method is to increase the number and diversity of road surface disease images by geometric transformation.The transfer learning method is to pre-train Faster R-CNN using ImageNet data set,and on the basis of the pre-training model,the parameters are optimized with the data set of road surface disease,which is finally used in road surface disease detection.These two methods improve the robustness of the model,reduce the over-fitting and improve the detection accuracy.