Enhancing Few-shot Learning for Metal Defect Classification with CBAM and Graph Embedding

Manual defect detection methods are constrained by challenges like operator fatigue and the requirement for specialized knowledge, which can lead to inaccuracies and suboptimal performance. To address these issues, this research presents an automated detection technique that capitalizes on the latest developments in deep learning (DL) and computer vision (CV) technologies. It specifically targets the problem of classifying metal surface defects with few training examples. A significant contribution of this work is the incorporation of the Convolutional Block Attention Module (CBAM), which substantially boosts the expressive capacity of the extracted features. This is followed by the application of Graph Embedding (GE) and Optimal Transport (OPT) methods to further refine the classification accuracy. The CBAM component enables the model to concentrate on salient image features by employing attention mechanisms across both channel and spatial aspects. Empirical results show that our approach achieves a 1.76<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> enhancement in accuracy for one-shot classification and a 5.00% increase for five-shot classification when compared to the conventional Graph Embedding and Optimal Transport (GEOT) techniques. Furthermore, we have developed a novel dataset, MSD-Cls, which includes 20 distinct categories of metal surface defects. This dataset serves as a benchmark to assess the efficacy of our proposed technique. The findings demonstrate that our method not only achieves higher classification precision but also better generalization across various scenarios, offering a robust solution for the classification of metal surface defects in industrial settings.