Improved faster R-CNN for steel surface defect detection in industrial quality control
Yuefeng Leng, Jiazhi Liu
Abstract
Steel surface defect detection constitutes a critical inspection task in industrial production. To address challenges including missed detections and low accuracy for fine defects, this study develops an enhanced Faster R-CNN algorithm. The proposed framework incorporates a feature fusion module and lightweight channel attention mechanism between Feature Pyramid Networks (FPN) and Region Proposal Network (RPN), substantially augmenting subtle feature extraction capabilities. Evaluated on the NEU-DET dataset, the optimized model achieves a mean average precision (mAP) of 80.2%-yielding a 12.6% improvement over the baseline-while increasing detection speed by 40.9%. This approach not only significantly elevates defect recognition accuracy but also establishes a practical framework for automated steel surface inspection systems.