deep learning, Prototypical Networks, Few-shot learning

This paper reports on the empirical performance of few-shot learning (FSL) for visualdefect classification using confidential industrial datasets. We evaluate 16 combinations offour backbone models (Perception Encoder, DINOv2, DINOv3, ConvNeXt-v2) and fourFSL classifiers (Prototypical Networks, Neighborhood Component Analysis, Relation Networks,Linear Adapter). The evaluation covers three conditions: a baseline comparison,deterministic support set augmentation, and a learnable attention preprocessor. Resultsdemonstrate that support set augmentation is a highly effective strategy, improving performancein nearly all configurations. Furthermore, the DINOv2 and ConvNeXt-V2-T backbonesemerged as top performers, achieving the most competitive and highest-accuracyresults, respectively. These findings suggest that for industrial FSL applications, combininga strong, pre-trained backbone with a simple augmentation strategy is a practicalapproach for building data-efficient classification systems.

This paper reports on the empirical performance of few-shot learning (FSL) for visualdefect classification using confidential industrial datasets. We evaluate 16 combinations offour backbone models (Perception Encoder, DINOv2, DINOv3, ConvNeXt-v2) and fourFSL classifiers (Prototypical Networks, Neighborhood Component Analysis, Relation Networks,Linear Adapter). The evaluation covers three conditions: a baseline comparison,deterministic support set augmentation, and a learnable attention preprocessor. Resultsdemonstrate that support set augmentation is a highly effective strategy, improving performancein nearly all configurations. Furthermore, the DINOv2 and ConvNeXt-V2-T backbonesemerged as top performers, achieving the most competitive and highest-accuracyresults, respectively. These findings suggest that for industrial FSL applications, combininga strong, pre-trained backbone with a simple augmentation strategy is a practicalapproach for building data-efficient classification systems.