A lightweight YOLOX-based model for detection of metallic gaskets on High-speed EMU wheelset

Abstract

Abstract In order to solve the requirements of real-time detection on metallic gaskets of wheelset assembly under the condition of limited hardware resources, here we propose an improved lightweight YOLOX-like network that can be deployed on mobile devices with low computing resource consumption. Firstly, the basic components of ShuffleNetV2 network structure are trimmed to reduce the computation intensity and the number of training parameters. Thus the lightweight ShuffleNetV2 is used as the backbone structure of YOLOX. The prediction layer for $$20\times 20$$ 20 × 20 size large targets is also removed from the head of this structure to further reduce the calculation time and improve the inference speed. Besides, the channel-attention enhancement module efficient channel attention is added into the network to improve the capability of feature extraction and the accuracy of target detection. Finally, the verification of inference for the proposed model is carried out on mobile terminal devices.The results show that the improved lightweight algorithm proposed in this paper not only ensures the detection accuracy, but also greatly reduces training parameters and computing resources, and it particularly can be run rapidly on mobile terminal devices with low-cost of computation.

Keywords

• High-speed EMU wheelset
• Bolt metal gasket
• Lightweight neural network
• Object detection