Online inspection of reflective metal surfaces requires accurate localization of tiny and low-contrast defects under strong background textures and specular highlights. Such conditions can weaken micro-defect cues after downsampling and lead to missed detections or background-induced false alarms. We present YOLOv11-LCCAP2, a deployment-oriented enhancement of YOLOv11 for steel surface micro-defect detection. It integrates two lightweight plug-in modifications: a stride-4 \(P2_{\textrm{small}}\) detection head to preserve high-resolution details for tiny defects, and a contrast-aware channel-coordinate attention module (LCCA) to strengthen weak defect responses while suppressing reflective background interference. We evaluate YOLOv11-LCCAP2 on NEU-DET and GC10-DET under a unified Ultralytics-based training and inference protocol, reporting accuracy and efficiency metrics including precision, recall, mAP, parameter count, GFLOPs, and end-to-end latency. To account for training variability under small performance gaps, we perform five-seed evaluation with paired statistical testing. Results show statistically supported improvements on NEU-DET, while gains on GC10-DET are modest and sensitive to initialization. Overall, the method improves micro-defect sensitivity with a clear accuracy–efficiency trade-off, and the reported end-to-end latency provides a practical reference for GPU-based online inspection.
