ASFC-Net: An Enhanced YOLOv8 for Heavy-Haul Rail Surface Defect Inspection via Asymmetric Spatial–Frequency Collaboration

Abstract

Weak rail-surface defects are difficult to detect because their localized evidence can be attenuated during downsampling and confused with structured rail texture. Across 2895 training and validation images, the low-, mid-, and high-frequency bands contain 97.23 %, 2.04 %, and 0.73 % of the whole-image spectral energy, respectively. This statistic characterizes global spectral imbalance but does not isolate defect-specific energy. We propose the asymmetric spatial–frequency collaborative network (ASFC-Net), a YOLOv8m-based detector that assigns different frequency-processing roles to P3 and P4. The low-loss semantic-guided high-frequency (LSGH) branch reconstructs localized frequency responses before further downsampling. Spatial–frequency centering alignment (SFCA) aligns centered spatial and frequency descriptors, while frequency-domain distribution-aware refinement (FDAR) combines learnable band priors with sample-dependent correction. Across five paired seeds, ASFC-Net achieves 60.38 % $\pm ~0.36$ % mAP@0.5 and 58.99 % $\pm ~0.49$ % recall, improving YOLOv8m by 0.99 and 1.11 percentage points, respectively; gap AP@0.5 improves by 2.23 points. The model uses 23.3-M parameters and 70.6 GFLOPs. The gains are concentrated at IoU 0.5: mean mAP@0.5:0.95 remains nearly unchanged, and batch-1 FP32 latency is higher despite the lower arithmetic complexity. These results support stage-asymmetric spatial–frequency processing for weak-defect detection while identifying remaining localization and runtime tradeoffs.