<p>Spatial-frequency-domain coupling is increasingly used in industrial surface defect detection. Leveraging frequency priors, recent models separate low- and high-frequency bands to sharpen contours and details, improving detection of tiny, extreme-aspect-ratio, and multi-scale defects. However, many methods rely on late concatenation or shallow injection, lack consistency-constrained alignment, and rarely distinguish the roles of low and high frequencies, hindering detail completion and precise localization. We propose SFCW-Net, a spatial-frequency-domain coupling wavelet network that addresses these issues. The Spatial Wavelet Block (SWB) couples a spatial branch with a wavelet branch: the spatial branch maintains global coherence, while the wavelet branch recovers localized high-frequency cues, making tiny defects easier to distinguish. The Wavelet-aware Multi-scale Modulation (WMSM) module uses three parallel paths. These include low-frequency subbands for channel calibration, high-frequency subbands for orientation-aware gains, and multi-scale depthwise convolutions for contextual modeling. Together, they jointly regulate channels, spatial context, and orientation, strengthening discrimination for extreme-aspect-ratio defects. The Dual-Path Token Fusion (DPTF) module improves cross-level alignment and selective aggregation; together with SWB, it forms a Token-Wavelet Feature Pyramid Network (TWFPN) that preserves fine details while maintaining scale alignment, reducing detail loss without sacrificing localization and improving recognition of multi-scale defects. The Local-Density and Geometry-Adaptive NMS (LDG-NMS) adjusts score decay according to local instance density and object geometry, mitigating both over-suppression and under-suppression and retaining small, elongated, and multi-scale instances more consistently.</p>