Lightweight and efficient railway foreign intrusion detection with feature enhancement and optimization

Abstract Achieving fast and effective detection of foreign object intrusion on railways is crucial for ensuring railway safety. Traditional detection models face challenges such as slow inference speed, complex structures, and difficulty balancing accuracy with efficiency. This study aims to design a lightweight and efficient model for railway foreign object detection, improving inference speed while maintaining high accuracy. First, we introduce the multi-channel enhanced feature attention net, a multi-channel feature enhancement network that improves multi-scale detection by enhancing critical spatial and channel features. Second, we propose the efficient shared head, a lightweight shared convolutional detection head that reduces parameters while maintaining localization and classification performance. Finally, we apply layer-adaptive magnitude-based pruning (LAMP) and Chrism knowledge distillation to reduce model size and improve performance. These methods reduce model parameters and memory usage without compromising detection accuracy. The method is evaluated on both the railway foreign object intrusion dataset and the public RailFOD23 dataset. On the railway dataset, the method achieves 88.7% mAP@50 and 67.7% mAP@50–95. LAMP and Chrism distillation reduce model size, and tests on edge devices demonstrate strong detection performance. Compared to state-of-the-art detectors, our model achieves higher accuracy, fewer parameters, and faster inference, demonstrating its superior performance and practical value.