Enhanced Cognitive Quality Inspection in Manufacturing through Siamese Few-Shot Learning

Defect detection in manufacturing remains challenging, with traditional methods relying on inflexible, hardcoded image processing techniques. While deep learning approaches show promise, they often lack scalability across product variations and require extensive training data. This paper presents a novel few-shot learning-based strategy for defect detection, addressing these limitations. Our proposed solution centers on a Siamese Defect classification model utilizing the Swin-L (Shifted Window Transformer) architecture. This approach enables efficient scaling across new products with minimal technical overhead. The models are trained on the industry-standard MVTEC dataset and optimized for real-time processing on embedded GPU devices, ensuring practical applicability in manufacturing environments. The developed prototype incorporates automated removal of defective items while preserving detailed information for subsequent analysis. By combining few-shot learning with advanced model architectures, our solution offers a reliable and highly scalable approach to defect detection.