Segmentation and Classification of Welding Beads on Shear Reinforcement With Dual Anchorage Using Deep Learning Models

ABSTRACT Automated visual inspection in industrial manufacturing requires robust and efficient instance segmentation systems capable of real‐time operation without compromising accuracy. This paper presents an enhanced 4‐camera framework for welding bead detection and quality assessment, leveraging a novel integration of optimized deep learning architectures and a Green's theorem‐based classifier. We conduct a comparative analysis of seven (7) state‐of‐the‐art models for instance segmentation, which are YOLO11, DeepLabV3+, Mask R‐CNN, YOLACT++, Cascade R‐CNN, SOLOv2, and a Hybrid CNN‐Transformer by evaluating their performance on a meticulously annotated dataset acquired under varying illumination and resolution conditions. The modified YOLO11 model is the main contribution of this work, particularly it unifies instance segmentation and classification within a single‐stage pipeline for a single class and object, achieving a superior balance between speed and accuracy. Experimental results show that while Cascade R‐CNN provides the competitive average precision (mIoU: 0.955 @ 0.5–0.95) and YOLACT++ attaining the best detection metrics (Accuracy: 0.974, Recall: 0.964), the proposed algorithm offers a competitive performance (Accuracy: 0.958, Precision: 0.935, Recall: 0.946) with an extraordinary speed exceeding 32 FPS. The system combines a four‐camera system and a mathematical‐based quality assessment rule that rules out beads derived from Green's theorem. This study shows that the suggested tool offers a practical, effective, and extremely accurate solution for welding bead segmentation and classification processes in real‐time, successfully meeting the crucial need for automated and dependable quality control systems in weld environments or industries.