MTC-YOLO: a multi-scale cross-layer perceptual approach for steel surface defect detection

Abstract Surface defect detection is crucial for ensuring the quality and safety of steel products. However, variations in defect scale, response intensity, spatial continuity, and morphology remain challenging. Conventional downsampling may lose fine-grained structures and weaken small defects; repeated spatial compression can attenuate deep target responses; and direct cross-layer fusion often overlooks heterogeneous defect patterns, causing feature mismatch and ineffective interaction. To address these issues, this paper proposes MTC-YOLO, a steel surface defect detection framework with multi-scale perception guidance and cross-layer enhancement. First, MESPD-Conv combines space-to-depth rearrangement with a non-uniform multi-scale receptive-field hierarchy and dual-attention recalibration to preserve fine-grained structures while reducing channel redundancy. Second, the target perception guidance module (TPGM) exploits local response variance to identify spatially consistent target regions and guide the exploration of potentially missing cues, enhancing the completeness of deep representations. Finally, the cross-layer enhancement module (CLEM) models sparse clustered and continuous defect patterns through spatial-semantic complementary interaction. Structure-aware aggregation, directional propagation, and dynamic cross-level weighting further improve hierarchical feature fusion while reducing feature mismatch. Experimental results on the NEU-DET and GC10-DET datasets demonstrate that MTC-YOLO maintains low parameter and computational costs while outperforming state-of-the-art methods (see e.g., CFC-CenterNet, ARYOLO, MPA-YOLO, STFE-Net, and CBH-YOLO) in mean Average Precision, with satisfactory robustness and applicability across different steel surface defect scenarios.