Enhancing object detection using YOLO-based model for weld quality identification

Abstract

This research developed a YOLOv10S model to detect and localize four types of weld quality: Crack, Porosity, Spatter, and Qualified welds. The study used 11,973 images from four public Roboflow datasets, divided into training, validation, and testing sets at an 80:10:10 ratio. The experiments investigated data augmentation, a weighted dataloader, loss functions, detection-head modification, and attention mechanisms. The results showed that basic augmentation combined with image resizing and the Focal-EIoU loss function improved the detection of small defects. while, without attention, the original detection head performed better than adding a P2 detection head. Installing CBAM at the P3 Upsample position achieved 82.51% precision and 82.44% mAP@0.50, which was 2.01% higher than the baseline model. It also achieved 60.95% mAP@0.50:95, which was 8.89% higher than the baseline model. The study further found that using a single CBAM module was more effective than combining it with Coordinate Attention.Therefore, the best-performing model was YOLOv10S with data augmentation and image resizing, the Focal-EIoU loss function, a four-head detector that includes the P2 tiny-object head, and CBAM installed at the P3 Upsample position.