Research On Rail Damage Detection Based On Faster R-CNN

China’s railway transportation is in a stage of rapid development,and the rail damage detection is an important part of railway safety.There are many methods for detecting rail damage,but it is still difficult to meet the requirements of accuracy and speed.Convolutional neural network is a special neural network model that can process two-dimensional matrix data.It is a new type of multi-layer artificial neural network generated by combining deep learning and artificial neural network.Non-full connection and weight sharing can reduce the computational complexity and the number of weights in the network training process,making the convolutional neural network highly variable in translation,rotation,tilt,scaling,etc.R-CNN,Fast R-CNN and Faster R-CNN network structure model based on convolutional neural networks are the development of convoluted neural networks.Based on deep learning and the convolutional neural network as the core,this paper designs a rail damage detection algorithm based on Faster R-CNN,which can automatically generate image feature problems from a set of rail damage and non-injury ultrasonic image training.The feature extractor eliminates the need to manually input features and damage area information to achieve the purpose of identifying rail damage.At the same time,this paper optimizes and improves Faster R-CNN from the aspects of data annotation,migration learning initialization model,parameter setting and training method.The test results show that the proposed method good at detecting on rail damage detection.By comparing and analyzing the experimental data of detection accuracy and detection time,the improved Faster R-CNN algorithm has better detection performance than traditional image processing and Faster R-CNN network model.The Faster-RCNN network model used in this paper uses the Region Proposal Network(RPN)algorithm,which is faster than the SS(Selective Search)algorithm when selecting candidate regions in the convolutional neural network model.At the same time,the improved network recognition accuracy is higher than the anchor boxes setting without improvement,so that the small-sized damage in the rail damage picture has a good detection effect.The experimental results show that the proposed recognition rate is 92.3%,which is better than the traditional network model based detection method.It is proved that the accuracy of the detection of the size and type of rail damage is better,and it has reached the main purpose of this paper.It has certain application value in the field of rail damage detection.Figure [30] table [6] reference [52]...