Unified Framework for MLCC Defect Detection Using Deep Learning Model Ensemble

The multi-layer ceramic capacitor (MLCC) is a critical electronic component that requires precise defect detection. However, detecting defects in MLCC cross-sectional images is challenging due to their high resolution and the subtle nature of defects such as slight distortions and misalignments. This paper presents a deep-learning-based framework integrating semantic segmentation, image classification, and anomaly detection for reliable defect identification. Key solutions include margin ratio threshold adjustment, contrastive learning, and a pseudo-defect memory bank. Experimental results demonstrate that our framework achieves 100% accuracy by leveraging the complementary strengths of individual models, with Soft-SupCon-based classification achieving a 96.82% micro F1 score, segmentation reaching a 92.71% mean intersection over union, and anomaly detection providing a 96.29% area under the receiver operating characteristic curve. These findings highlight the potential of deep learning for MLCC defect inspection and its applicability to industrial settings.