| Rail as an important part of railway,under the action of train extrusion and external environment for a long time,it is easy to appear some injuries,these injuries are mainly manifested as rail surface injuries.These rail surface damage will pose a major hidden danger to the safe operation of trains,and even lead to derailment,capsizing and other accidents.Compared with the traditional rail surface damage detection method,the deep learning-based rail surface damage detection method has higher detection accuracy and faster detection speed.However,the current rail surface damage detection algorithm based on deep learning still can not meet the requirements of the rail surface damage detection task on the detection speed and accuracy of the algorithm.To solve the above problems,an algorithm based on YOLOv4 rail surface damage detection is proposed in this thesis.The YOLOv4 algorithm has the disadvantage of slow running speed of the overall model in rail surface damage detection.In this thesis,Mobilenetv3 lightweight feature extraction network based on SA attention mechanism and Shufflenetv2 lightweight feature extraction network based on SA attention mechanism are proposed,and then Mobilenet-SA-Yolov4 network and ShufflenetSA-Yolov4 network are constructed.The experimental results show that the detection speed of Mobilenet-SA-Yolov4 network and Shufflenet-SA-Yolov4 network is increased to 53 FPS and 62 FPS respectively under the premise of certain detection accuracy.In addition,Shufflenet-SA-Yolov4 network is superior to Mobilenet-SA-Yolov4 network in the overall model.In terms of the loss function,the penalty term of the size of the predicted box and the real box is added to the loss function to make up for the poor positioning of the prediction box when the shape of the predicted box and the real box is similar.The experimental results show that the detection accuracy of the algorithm is improved by 0.56% by this method.In view of the poor detection accuracy of small targets in rail surface damage detection by YOLOv4 algorithm,the structure of YOLOv4 network was improved in this thesis.The bidirectional weighted feature fusion network is used to replace the original feature fusion network,and the second-layer features of the backbone network are introduced to improve the detection accuracy of the overall algorithm for small targets.Aiming at the shortcoming of coupling between classification task and regression task in YOLOv4 algorithm,the decoupling method is used to decouple the output header of the algorithm,which further increases the detection accuracy of the algorithm.Experimental results show that the above two methods increase the detection accuracy of the whole algorithm in different degrees. |
