Comparative Study of YOLO Architectures for Detecting Complex Weld Defects under Label Ambiguity

Abstract

The objectives of this research are 1) to develop a weld defect detection model utilizing the YOLO architecture, specifically three versions 2) to compare the performance of the YOLOv11l, YOLOv12l, and YOLOv26l models under identical experimental conditions; 3) to analyze the class-specific weld defect detection results, which include Crack, Porosity, Spatter, and Weld Line; and 4) to develop a web application prototype for weld defect detection. The study was conducted by examining a dataset of 5,544 annotated images comprising four defect categories: Crack, Porosity, Spatter, and Weld Line. All models were trained with identical configurations and assessed. The results indicate that YOLOv26l achieves the highest overall performance, particularly at 300 epochs, yielding a Precision of 0.91, Recall of 0.876, mAP50 of 0.908, mAP50–95 of 0.773 and F1-score of 0.89. These metrics reflect superior localization accuracy and defect classification capability compared to the other architectures. Class‑wise analysis reveals that Crack and Weld Line are consistently well‑detected, whereas Porosity and Spatter remain more challenging due to visual similarity with background textures. Despite this, YOLOv26l demonstrates the best balance across all defect classes. Furthermore, the trained model was successfully deployed in a web‑based application, enabling prototype demonstration for weld‑quality inspection. The findings confirm the potential of modern YOLO architectures, especially YOLOv26l, for accurate and robust automated weld defect detection in industrial environments. However, this research is limited by the use of only a single public dataset, and therefore may not cover all real-world environments.