LLRailDet: Task-Specific Asymmetrical Decoupling for Robust Low-Light Railway Defect Detection

Railway inspection in low-light conditions is critical for safety but remains challenging due to degraded visibility, noise, and complex background interference. While one-stage detectors are efficient, their symmetrical decoupled heads, which use identical structures for classification and localization, are suboptimal in this domain. This limitation arises because classification relies on discriminative features, whereas localization requires precise spatial and edge information. This paper introduces LLRailDet, a one-stage detector designed to address this limitation. Our model employs an existing lightweight Transformer backbone (LVT) for multi-scale feature extraction and a Feature-Mix Neck (FMN) that combines spatial-channel attention with frequency-domain processing to suppress noise and enhance edges. Crucially, we propose Asymmetrical Decoupled Heads (ADH) tailored to the distinct objectives of each task: a Multi-Axis Fusion Module (MAFM) for classification, which aggregates global context to improve feature discrimination, and a Group Channel Aggregation Module (GCAM) for regression, which refines spatial details through progressive channel-wise fusion. Extensive experiments on a collected low-light railway dataset demonstrate that LLRailDet achieves a significant performance gain, outperforming YOLOv10 by 6.5% AP and Dab-DETR by 0.4% AP while maintaining real-time efficiency. On converted low-light defect datasets, such as NEU Det, DAGM2007, and Magnetic Tile, our model also demonstrates superior generalization.