Deep Learning-based Recognition of Welding Defects in Spiral Submerged Arc Welding Images

This study explores the use of the YOLOv5-master model for defect detection in images of spiral submerged arc weld seams. Given the pivotal role of image preprocessing in optimizing deep learning model performance, this work highlights the impact of these preprocessing steps on enhancing model learning and training efficiency. Various preprocessing techniques—including standardization, normalization, cropping, rotation, scaling, and flipping—were applied to ensure data quality and consistency, leading to improved model generalization and detection accuracy. Experimental results demonstrate that a well-structured preprocessing workflow significantly boosted the YOLOv5-master model’s performance in identifying welding seam defects.