arXiv.itsnick.co

an independent field guide to 23 million papers from arXiv & OpenAlex — not affiliated with arXiv.org

Computer Science · Artificial Intelligence

Domain-Adaptive Prototype Networks for few-shot intangible cultural heritage image classification

Yan Nie, Guanghao Xu, Jing Qiu, Yunqian Wang

· Alexandria Engineering Journal

See on the map

Abstract

The digitization of intangible cultural heritage (ICH) often proceeds through the cataloging of physical craft artifacts that preserve visual traces of techniques, materials, and regional styles. This artifact-centered setting is challenging because labeled examples are scarce and support-query pairs frequently differ in region, material aging, illumination, and capture scale. We propose Domain-Adaptive Prototype Networks (DAPN), a meta-learning framework that combines invariant prototype learning, feature disentanglement, and graph-based prototype refinement for few-shot artifact image classification under domain shift. DAPN is designed to preserve morphology and style cues that remain stable across domains while isolating material- and region-dependent variation in a separate branch. Instead of claiming a formal proof of tighter bounds, we provide a domain-adaptation-based analysis that explains why reducing divergence in the invariant subspace should improve prototype transfer, and we state the assumptions under which this interpretation is reasonable. Experiments on ICH-Benchmark, mini-ImageNet, and tiered-ImageNet show that DAPN improves on the strongest baseline by 1.3–1.8% and exceeds Prototypical Networks and MAML by 11.6% and 14.3%, respectively, in the 5-way 1-shot setting on ICH-Benchmark. These results support DAPN as a practical tool for artifact-centered ICH documentation when labeled data are limited.

More like this