Research on Intelligent Operation and Maintenance of Rail Transit Based on Improved DCGAN Data Enhancement

This paper proposes a rail fastener detection method combining improved DCGAN and improved YOLOv5 to improve the detection accuracy of rail transit fastener. On the one hand, the method expands the sample size by DCGAN, and optimizes the network structure by adding deconvolutional layers, so as to improve the recognition effect as a whole. On the other hand, the method replaces the YOLOv5 backbone network with GhostNet and adds the CA attention mechanism, so as to make YOLOv5 can retain the detection accuracy on the basis of lightweight. The results show that improving DCGAN can expand sample data, and the SSIM value is 0.9756 when the training reaches 15000 times. By analyzing the detection results of the same detection network on the improved DCGAN expansion data set, it can be seen that the mPA is increased by 28.65% and 25.92% compared with the original sample and the traditional data enhancement method. By adopting the proposed improved YOLOv5 track fastener detection method on the same data set, the mPA is increased by 11.65%, 19.83% and 4.52% compared with the current mainstream detection methods YOLOV5-Tiny, Faster-RCNN and Shuffle R-CNN. Although it is slightly lower than YOLOv5, the memory is reduced by 90.39% compared with YOLOv5, showing the best overall performance. It has higher practical value.