APFNet: A lightweight attention-based adaptive perception and fusion network for no-service rail surface defect segmentation

The safe operation of railway systems critically depends on rails, for which surface quality detection plays a crucial role in ensuring reliability and safety. Recently, vision-based approaches have demonstrated clear advantages in no-service rail surface defect (NRSD) detection. However, lightweight single-modal methods specifically designed for NRSD segmentation remain unexplored, which are well suited for deployment on resource-constrained industrial inspection devices and thus possess greater practical applicability. To bridge this gap, this paper proposes an attention-based adaptive perception and fusion network (APFNet), which is a lightweight yet effective segmentation framework specifically tailored for NRSD scenarios. APFNet is motivated by an in-depth analysis of NRSD characteristics and targets the fundamental difficulties inherent to NRSD segmentation. Specifically, a saliency-aware feature decoupling module is introduced to enhance defect saliency and suppress background interference under low-contrast conditions. An attention-based adaptive perception and fusion module is designed to capture multi-scale contextual information, enabling robust modeling of randomly distributed defects with diverse scales. Furthermore, an edge-assisted collaborative decoding module refines defect representations by jointly preserving edge integrity and regional consistency, leading to more accurate delineation of irregular defect morphologies. Extensive experiments conducted on four public datasets demonstrate that APFNet consistently outperforms 17 state-of-the-art methods while maintaining a compact architecture, with only 1.54 M parameters and 1.21 G FLOPs. These results indicate that APFNet achieves a favorable balance between accuracy and efficiency, making it well suited for practical deployment in real-world industrial inspection systems. The code will be available at https://github.com/Qian-Cheng-Li/APFNet.