Few-Shot Welding Defect Detection via Graph Self-Attention Neural Network

Abstract

In industrial quality control, deep learning-based weld defect detection has shown promise but is often hindered by the need for extensive labeled datasets. This study introduces a novel graph neural network (GNN) approach for few-shot classification of weld defect images, aimed at mitigating data dependence and enabling efficient defect detection in data-scarce environments. The goal is to achieve high accuracy with minimal labeled samples, reducing data collection costs and enabling applications in scenarios with limited annotated data. A GNN-based algorithm is designed that incorporates a patch mechanism to reduce computational complexity and a self-attention mechanism to enhance node feature representation. The model is evaluated on public datasets (Al5083 and SS304) using only 5% of the training data, and its performance is compared against the ResNet18 baseline. The proposed model achieves accuracies of 82.0% on Al5083 and 94.8% on SS304, outperforming ResNet18 by 5.2% and 2.4%, respectively. Results demonstrate strong generalization performance with limited training samples, validating the efficacy of the approach in few-shot settings. This work reduces dependency on large, annotated datasets, lowering the cost and effort of data collection in industrial environments. It provides a practical solution for high-precision detection of weld defects in data-scarce scenarios, with potential applications in specialized manufacturing contexts.