A New Lightweight Network for Real-time Detection of Railway Infrastructure

The deployment of deep convolutional neural network in railway infrastructure real-time detection is largely hindered by huge computing costs. This paper proposes a new lightweight network called S-CA-YOLOv5, in which a new backbone combined ShuffleNet with Coordinate Attention is used to reduce the computational complexity, and the CA mechanism can capture regions of interest accurately. Moreover, the improved loss function with Efficient Intersection over Union facilitates more accurate localization and classification, which is conducive to higher detection precision and faster convergence speed. Finally, through channel pruning combined with knowledge distillation fine-tuning, the model is further lightweight. Compared with other methods, our proposed method achieves the best trade-off between detection precision and speed. Hence, the proposed method has the potential to achieve real-time detection for the railway infrastructure based on embedded Graphic Processing Unit platforms.