Embedded vision transformers for industrial welding defect detection

Detecting defects in Tungsten Inert Gas (TIG) aluminium welding is crucial for industries such as marine, aerospace and automotive, where weld quality impacts safety and performance. This study investigates the performance of Vision Transformer (ViT) models for detecting welding defects in the aluminium 5083 TIG dataset. In addition, this study evaluates ViTs on embedded Graphics Processing Unit (GPU) boards using benchmarks for resource-constrained devices. Although many studies have explored Machine Learning (ML) or Deep Learning (DL) techniques for welding defect detection, they often overlook performance evaluation metrics, such as inference speed, on embedded hardware. We provide a comparative analysis of ViT models, including Swin Transformers and traditional ViTs, evaluating defect detection performance using metrics like F1-score, precision, recall, accuracy, Frames-Per-Second (FPS), and memory consumption on embedded GPU hardware. Results from the TIG Aluminium 5083 dataset show that the DeiT-Base model achieved an accuracy of 99.83%, outperforming existing models. These results confirm the effectiveness of ViTs for automated welding defect detection on both conventional computers and resource-constrained embedded systems. Finally, we successfully implemented ViT-based models on embedded hardware, showing robust performance even under limited computational resources.