Deep learning-based detection transformer (DETR) for defect identification in shielded metal arc welding plate joints

Traditional inspection methods for Shielded Metal Arc Welding (SMAW), including visual inspection and radiographic testing, are labor-intensive, time-consuming, and prone to subjectivity. As industries move toward automation and real-time quality assurance, deep learning-based computer vision presents a compelling alternative. This study investigates the application of the Detection Transformer (DETR) with a ResNet-50 backbone for automated detection of surface-level SMAW defects, specifically spatter, slag inclusion, and undercut. A robust dataset of annotated high-resolution weld images was developed using SMAW on SS45 steel, with images collected from diverse facilities and augmented to improve generalizability. The DETR model was trained and evaluated using five-fold cross-validation, achieving stable performance with mAP@0.5 exceeding 0.76. However, detection of subtle defects like undercut showed lower performance at stricter IoU thresholds (mAP@[0.5:0.95] ≈ 0.56). Expert validation yielded 71% accuracy under realistic image-level evaluation, confirming the model’s practicality. To enhance field applicability, the trained model was deployed via a REST API and integrated into an Android application, enabling mobile-based defect detection. This work contributes toward bridging the gap between deep learning research and real-world welding inspection, demonstrating that transformer-based detectors like DETR offer both accuracy and deployability for industrial defect detection workflows.