Confidence–gradient reweighting and lightweight feature enhancement algorithm for steel surface defect detection

Abstract

Abstract Steel surface defect detection is susceptible to small target sizes, low contrast, and class imbalance. To this end, we propose the Gradient-Reweighting with Awareness of Confidence and Lightweight Feature Enhancement (GRACE) algorithm built upon YOLO11s, composed of two synergistic modules: Dynamic Sampling with Confidence-Gradient Balanced Sampling Mechanism (DS-CBSM++) performs dynamic reweighting via joint confidence-gradient feedback, improving the separability of hard examples and long-tailed classes; Lightweight Feature Enhancement Network (Lite-FEN) introduces lightweight channel/spatial enhancement at the P3 layer to strengthen shallow textures and boundary cues while keeping computation low. Experiments on the NEU-DET dataset show that the baseline YOLO11s achieves an mAP@0.5:0.95 of 42.66% and an mAP@0.5 of 74.69%. GRACE achieves 43.66% and 75.88%, respectively, improving over the baseline by 1.00 percentage points and 1.19 percentage points, with 9.56 M parameters, suitable for real-time inference. These results indicate that GRACE yields more robust detection and localization of small defects under complex textured backgrounds.Additional experiments on the GC10-DET and X-SDD datasets further confirm that GRACE maintains competitive performance across different steel surface defect distributions.