Ancient glass beads excavated across Japan are key materials for understanding cultural exchanges across the Eurasian continent. Conventional analysis relies on expert inspection and chemical composition measurements, making large-scale investigations difficult. To address this issue, we propose a machine learning framework for classifying ancient artifacts using only image data. To correctly classify ancient glass beads based on their provenance, capturing both fine surface details and overall shapes is essential. We achieve this by integrating convolutional neural networks (CNNs) and vision transformers (ViTs) into a single architecture and employing dual-view (Top and Side) image fusion to emulate expert observation. Furthermore, our framework incorporates bidirectional knowledge distillation between models and viewpoints to ensure feature consistency and efficient multiview fusion. Experiments demonstrate that although the proposed method does not always achieve the highest average accuracy, it exhibits the lowest performance variance across folds, indicating stable classification performance even for rare classes. This study demonstrates for the first time that an image-only approach is effective for the provenance estimation of ancient glass beads. Our findings demonstrate the potential of image-based approaches for analyzing archaeological materials and highlight their promise for advancing archaeological research.
FUKUCHI et al. (Sat,) studied this question.