A hybrid Swin–DeiT transformer framework for automated welding defect detection in radiographic images

• Swin Transformer backbone integrated with DeiT classification head. • Image-level split with 224 × 224 patches prevents data leakage. • SWRD dataset (3675 X-ray images, 7 defect classes) utilized. • Achieved 99.55% binary and 98.36% multi-class accuracy. • Ablation, cross-validation, and complexity analysis validate model. . Weld defect detection involves identifying issues such as pores, lack of fusion, and cracks in welded joints to ensure structures are safe and strong. In industries, using image classification to automatically detect these problems helps to reduce human errors, improve quality checks, and increase accuracy. It is necessary to improve safety, reliability, and productivity in automatic welding inspection systems. This research aims to build a smart and reliable system for detecting weld defects using image classification. This study uses deep learning models called Swin Transformer and DeiT (Data-efficient Image Transformer). The research uses the SWRD (Seam Weld for Defect Detection) dataset, which has images of major weld defect types. The aim is to obtain the best results in both multi-class and binary classification. To make the images clearer and better for training, preprocessing steps like image denoising and image enhancement are applied. Tests were done using the SWRD dataset for both multi-class and binary tasks. In binary classification, the model achieved 99.55% accuracy, 99.43% precision, 99.29% recall, and a 99.38% F1-score. In multi-class classification, it reached 98.36% accuracy, 98.4% precision, 98.3% recall, and a 98.3% F1-score. These results demonstrate the efficacy of the developed method for identifying common weld defects.