Enhanced Deep Learning Framework for Steel Surface Defect Detection under Class Imbalance

Abstract

Steel surface defect detection is a critical task in intelligent manufacturing, where accuracy and real-time performance are essential for ensuring product quality. Although recent deep learning–based object detection methods have achieved promising results, challenges remain in handling complex defect morphologies, scale variations, and class imbalance. To address these issues, this paper proposes an enhanced steel surface defect detection framework based on a lightweight one-stage detection architecture. Specifically, Deformable Convolution v3 (DCNv3) is introduced into the detection head to improve the model’s adaptability to geometric deformations and irregular defect patterns. In addition, Focal Loss is employed to alleviate the imbalance between foreground and background samples, enabling the model to focus more effectively on hard-to-detect defects. The proposed method is evaluated on the NEU steel surface defect dataset under a consistent experimental setup. Experimental results demonstrate that the proposed approach achieves notable improvements in detection accuracy while maintaining real-time performance, outperforming the baseline detection framework in terms of mAP_0.5. These results indicate that the proposed framework is effective and practical for industrial surface defect inspection tasks.