Defect detection in mass production lines is often challenged by small and imbalanced datasets, making few-shot learning approaches particularly suitable. Traditional deep learning methods typically rely on large-scale datasets for training, which limit their applicability in real-world manufacturing environments. To address this limitation, this study proposes a few-shot learning model capable of identifying product defects using a limited amount of data, thereby enhancing generalization across multiple defect classes. Unlike conventional deep learning models that require extensive data, the proposed approach effectively performs defect detection with minimal samples. Specifically, we introduce a Siamese Convolutional Neural Network-based Few-Shot Learning (SCNN-FSL) framework. The Siamese network is constructed using CNN architecture and trained with a triplet loss function to optimize feature embedding. Furthermore, SCNN-FSL is integrated into an automated optical inspection (AOI) defect detection system, incorporating image preprocessing, data sampling, and object classification techniques tailored for detecting defects in electronic components on PCB production lines. Experimental results demonstrate that the proposed few-shot learning model outperforms traditional deep learning approaches, achieving higher accuracy and lower miss rates, thereby validating its effectiveness in practical industrial applications.
關聯:
2025 IEEE 14th Global Conference on Consumer Electronics (GCCE )
