OAD-YOLOv8n: A Lightweight Direction-Adaptive Framework for Steel Strip Surface Defect Detection

Steel strip surface defect detection remains challenging because defects are often elongated, weakly bounded, low-contrast, and sensitive to imaging degradation. To address these issues, this paper proposes Orthogonal Direction-Adaptive YOLOv8n (OAD-YOLOv8n), a lightweight detector based on You Only Look Once version 8 nano (YOLOv8n) and centered on Orthogonal Direction-Adaptive Efficient Multi-Scale Attention (OA-EMA), an orthogonal direction-adaptive attention module that combines debiased strip descriptors, adaptive direction selection, and local directional convolution. Dynamic upsampling by learning to sample (DySample), a lightweight neck structure (SlimNeck), and Adaptive Threshold Focal Loss (ATFL) are further integrated to improve detail-preserving upsampling, efficient multi-scale fusion, and hard-sample optimization. Across five independent runs on NEU-DET, OAD-YOLOv8n improves Precision, Recall, mAP50, and mAP50:95 by 5.0, 3.6, 4.4, and 3.7 percentage points over YOLOv8n, while reducing FLOPs and parameters by approximately 10.3% and 7.0%, respectively. Complementary experiments on GC10-DET, cross-dataset transfer/adaptation, simulated practical image perturbations, failure cases, and measured inference speed provide a broader characterization of the model’s benchmark-level generalization, robustness, and deployment-related behavior. These results indicate that OAD-YOLOv8n provides an effective accuracy–efficiency trade-off for lightweight steel strip surface defect detection.