WMS-Net: a multi-stage Mamba network with wavelet multi-component fusion for industrial defect detection

Industrial defect detection often struggles with small defects, weak boundaries, and distracting background textures. We present WMS-Net, a detector that combines wavelet decomposition, stage-wise feature injection, and multi-stage feature aggregation. The wavelet transform module separates structural information from directional details, while the injection framework sends these components to different network stages: low-frequency information first supports defect localization, followed by high-frequency components that refine edges and textures. A Mamba-based aggregation module then combines the injected details with spatial context. On the real production-line CDD-25k dataset, WMS-Net achieves 92.5% mAP50 and 71.4% mAP50-95. It also reaches 85.0% and 79.3% mAP50 on DAGM2007 and RDD2022, respectively. Additional ablations support the selected wavelet basis, component allocation, and injection order. Experiments across real production-line, textured-surface, and natural-scene datasets demonstrate the effectiveness and cross-scene applicability of the proposed frequency-aware detection framework.