SDCM-YOLO: a cross-scale, texture-aware framework for metal surface defect detection

Abstract Metal surface defect detection is often hampered by small defect sizes, low contrast, and strong background textures, causing conventional approaches to suffer from frequent misses and false alarms in real-world settings. To balance detection accuracy with model compactness, we develop SDCM-YOLO, an improved framework built on YOLO11. Methodologically, we incorporate Shallow Robust Feature Downsampling (SRFD) and Deep Robust Feature Downsampling (DRFD) into the backbone and neck to preserve edge and fine-texture contrast while stably propagating high-level semantics under a fixed compute budget. We further design a Cross-Stage Partial Feature Aggregation (CSPFA) block that combines grouped 3 × 3/5 × 5/7 × 7 convolutions with hierarchical aggregation, strengthening cross-scale information flow and fusion efficiency. In addition, we propose a novel MSW (Morph–Spectral–Window) fusion module at the same-resolution lateral connections in the neck to replace naive concatenation. MSW applies joint gating based on morphological gradients, spectral energy, and local-window consistency to modulate fused features in a controlled manner, suppressing the cumulative amplification of background streaks and specular highlights while enhancing defect edges and fine structures. Experiments on the AL-DET and NEU-DET datasets confirm the effectiveness of the proposed design: with ∼2.657M parameters and 8.9 GFLOPs, SDCM-YOLO attains mAP@0.5 scores of 75.2% and 77.0%, representing absolute gains of 7.6 and 3.8 percentage points over the baseline. The model thus achieves strong overall performance while remaining lightweight, indicating promising potential for practical deployment.