Real-Time Detection Of Track Fasteners Image Based On Deep Convolutional Neural Network

The health status of railway track fasteners is essential for the safe and stable operation of the train.At present,automatic inspection by trains instead of manual patrols has gradually become the main inspection manner,but the method that has been adopted still do not meet the requirements of high-speed and accurate inspection,whereas the research on the real-time detection method of track fasteners is conducive to improving the precision and speed of track inspection trains.The research including real-time detection method of track facilities and lightweight real-time detection system for fasteners has significantly practical significance for improving the efficiency of real-time inspection of China's track status.However,the task is also challenging due to the limited memory and processing capacity of the embedded maintenance systems.The current detection network has many shortcomings such as low detection accuracy,large memory occupation and detection speed that still need to be improved.In this paper,the detection network YOLOv3-Tiny is newly designed and applied to the real-time detection of fasteners,and a set of high-definition track fasteners data collected on site is constructed to test and evaluate the designed network.In this paper,the construction of a deep convolutional neural network model for track fasteners detection is focused,and main contributions of this paper are summarized as follows:(1)A new track fastener detection network architecture called MYOLOv3-Tiny is proposed.It combines a lightweight detection network YOLOv3-Tiny and a backbone with linear bottlenecks and inverted residuals to improve the detection precision and efficiency.(2)A hybrid convolution strategy is proposed,which integrates a depthwise separable convolution with the standard one.This strategy has the effect of drastically reducing computational complexity and model size,which makes the proposed method guarantee the real-time requirement of the embedded computer vision devices.(3)A method of detection combining the YOLOv3-Tiny and bottleneck depth-separable convolution with residual is adopted.The main advantage is to guarantee the detection precision by efficiently extracting the fasteners' features while significantly reducing memory consumption,which is especially helpful for the embedded fastener detection devices.(4)In addition,extensive experiments are conducted based on the real-world image data collected from the Ring Test Field of National Railway Test Center of China.The results show the proposed MYOLOv3-Tiny achieves higher detection precision,lower computational complexity and memory consumption,and faster detection speed compared to state-of-the-art methods.Hence,it has the potential to be applied to large-scale track detection devices.Combining the deep separable convolution and linear bottleneck structure,a new fastener detection network model called MYOLOv3-Tiny is proposed,the performance of network model is evaluated and verified based on the high-definition track fasteners image data set.Its detection precision reaches 98.90%,moreover memory consumption is reduced by 43% and has higher detection speed.The performance of network model MYOLOv3-Tiny to real-time detect and identify track fasteners gets validated on the basis of these results.The MYOLOv3-Tiny network model proposed in the paper has remarkable guiding significance for the precision,speed optimization and lightweight development of track fasteners real-time detection system,which has the potential to lay a foundation for the real-time detection of railway track fasteners and to be applied to large-scale fastener detection devices.