DBCW-YOLO: An advanced yolov5 framework for precision detection of surface defects in steel
The primary goal of this study is to enhance surface defect detection in steel manufacturing through advanced machine learning techniques. Traditional inspection methods often fall short in terms of accuracy and efficiency, necessitating more robust solutions. This paper presents DBCW-YOLO, an optimized version of the YOLOv5 framework, designed to improve defect detection by integrating attention mechanisms and enhanced feature extraction techniques. The model incorporates the BiFPN technique to maximize feature map fusion across multiple scales, and CARAFE upsampling expands the network’s receptive area, effectively utilizing neighboring data for better detection. Additionally, WIoU introduces a dynamic, non-monotonic focusing mechanism in the loss function, addressing the issue of accuracy degradation caused by sample inhomogeneity. The model's performance was evaluated on multiple defect categories: crazing (Cr), inclusions (In), patches (Pa), pits and scales (PS), roll scale (RS), and scratches (Sc). DBCW-YOLO demonstrated significant improvements in Average Precision (AP) compared to the baseline YOLOv5m model, with AP for Cr increasing from 35.69% to 70.25%, In from 60.25% to 76.85%, and PS from 59.74% to 92.36%. These results confirm that DBCW-YOLO is a highly effective and efficient tool for automating surface defect detection in the steel industry, significantly outperforming the baseline model.
The primary goal of this study is to enhance surface defect detection in steel manufacturing through advanced machine learning techniques. Traditional inspection methods often fall short in terms…
Purpose