Research On Track Fastener Detection Algorithm Based On Deep Learning

As an important part of railway track infrastructure,track fasteners are the key to the safe operation of trains on the track.At present,the track line fault inspection vehicle based on computer vision has replaced the traditional manual inspection as the main detection method,but its detection accuracy and speed cannot fully meet the requirements.Therefore,the research on the high-precision real-time detection method of track fasteners is helpful to improve the detection effect of track inspection vehicles.With the rapid development of deep learning theory,target detection algorithms based on deep convolutional neural networks have been widely used in various fields.In this paper,the target detection algorithm based on deep learning is applied to the detection of rail fasteners,mainly from the following aspects:(1)Aiming at the problem that there is no public data set for rail fastener detection,based on the analysis of the basic knowledge of public target detection data sets,this paper produces a rail fastener detection data set containing three categories of normal,damaged and missing fasteners.Annotation software was used to label the sample target information in the dataset.The dataset contains 11046 fastener image data and 47580 fastener sample information.(2)In view of the imbalance of positive and negative samples in fastener target detection,resulting in poor fastener detection effect,this paper optimizes on the basis of YOLOv3 detection algorithm,and introduces Focal Loss loss function into the loss function to replace the original YOLOv3 algorithm Loss function confidence cross entropy loss function,reduce the weight of negative samples in training;use GIOU Loss loss function as the target box coordinate regression positioning sum of squares loss function.The average accuracy of detection of small fastener objects is improved,resulting in faster convergence.Then upsampling is used in the detection layer to double the feature maps,and the feature maps of the same size are connected in the backbone network,and finally fasteners are detected at four scales,which improves the network model’s ability to detect fasteners.The experimental results show that the improved YOLOv3 algorithm achieves an average accuracy of 78.36%on the self-made fastener data set.Although it achieves high detection accuracy when detecting fasteners,the detection speed cannot meet the requirements of real-time detection.(3)In order to further improve the detection speed of the network model for fasteners and meet the real-time requirements of the application.This paper adopts the lightweight network model YOLOv3-tiny algorithm,the backbone network uses traditional convolution and combined convolution with a depthwise separable inverse residual structure to extract the input coarse-grained features of fasteners,using a combined volume with stride 2 The product replaces the max pooling layer for downsampling to reduce the dimensionality of the feature map and reduce the number of backbone network weight parameters.In addition,the depthwise separable convolution is used in the detection layer to extract the deep-level feature information of the fastener,and then the upsampling method is used to increase the dimension of the feature map,and the feature maps of the same size are connected in the backbone network,increasing a scale in detection on three scales.The loss function uses the DIOU Loss loss function as the target box coordinate regression positioning sum of squares loss function,and the Focal Loss loss function instead of the confidence cross-entropy loss function in the lightweight network model algorithm,and the category probability still uses the cross-entropy loss function.The improved lightweight model YOLOv3-tiny algorithm is trained and tested on the track fastener data set in this paper,and the comparison experiment of the lightweight model YOLOv3-tiny algorithm before and after improvement with other target detection algorithms is completed.The experimental results show that the average accuracy of the improved lightweight model YOLOv3-tiny algorithm on the self-made fastener dataset is higher than that of other algorithms,and the average accuracy reaches 85.69%.Although the detection speed has decreased,it meets the requirements of real-time detection of track fasteners.