Explainable Vision Transformer with ResNet18 Feature Extraction for Automated Welding Defect Classification using LIME

Abstract

Welding helps in the joining of two materials using high temperature and filler material. The material's complete properties depend on the welding quality, which makes the product more resilient towards the strong factors. It is common to encounter faults in welded materials; therefore, to address this issue, a visual inspection of the product is carried out. The development of Artificial Intelligence enables the automated detection of weld defects. Deep learning models are mainly used for defect detection, especially of CNNs. This paper proposes an explanatory model for the classification process of the Vision Transformer using ResNet18 feature extraction on the welding surface images dataset. The proposed work, using patches and attention, achieves efficient classification, and Local Interpretable Model Explanation is used to explain the features responsible for the defect and to rectify it. By this method, the proposed work showed 92.96% accuracy and 92.82% F1 score.