EF-Yolo: An Efficient and Lightweight Network for Real-Time Components Detection of Freight Trains

Component detection is crucial to the trouble of moving freight car detection system based on image recognition and analysis, serving as the foundational step for downstream fault analysis. However, existing detection algorithms struggle to meet stringent accuracy and speed requirements in real-world high-density vehicle dynamic inspection scenarios. Efficient (EF)-you only look once (YOLO) is proposed to enhance detection speed, whereas preserving high accuracy. EF-Yolo is a lightweight train component detection network that incorporates two efficient gradient-based feature extraction modules, where C2f-Ghost focuses on lightweighting and C2f-SE-Ghost focuses on feature extraction. Furthermore, we introduce a novel module deployment strategy that focuses on optimizing the arrangement of feature extraction modules based on the network’s architectural characteristics, thereby maximizing model inference efficiency and detection performance. Experimental results on a freight train components dataset demonstrate that EF-YOLO outperforms state-of-the-art detectors in competitive accuracy and achieves a remarkable detection speed of 70 frames/s, whereas maintaining a modest model size of only 9.15 MB, and can meet the efficiency requirement of completing the detection of 15000 images of a single train within ten minutes.