Enhancing Steel Surface Defect Detection: A YOLOv8-DCNv3-FOCAL Integrated Approach

Abstract

In the context of Industry 4.0, efficient and accurate surface defect detection in steel manufacturing has become increasingly critical. Traditional manual inspection methods suffer from limitations such as low efficiency and inconsistency, while machine vision-based techniques encounter high costs and sensitivity to environmental factors. Recent advances in deep learning, particularly the YOLO series algorithms, have introduced promising solutions for defect detection. This paper presents an improved approach based on the YOLOv8 model, incorporating DCNv3 and Focal Loss to address challenges in detection accuracy and real-time performance. The proposed YOLOv8DCNv3-FOCAL algorithm demonstrated a significant enhancement in mean Average Precision (mAP_0.5) value, reaching 81.6% during experimental validation. This improvement highlights the model's superior capabilities in capturing complex defects and mitigating classification imbalances. Despite these advancements, early-stage training instability remains a concern, which will be the focus of future research.