Domain-Adaptive Prototype Networks for few-shot intangible cultural heritage image classification
Year: 2026
Type: article
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.
Source: Alexandria Engineering Journal
Authors: Yan Nie, Guanghao Xu, Jing Qiu, Yunqian Wang
Institutions: Jiangxi Normal University, Nanchang University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,900
OpenAlex ID: W7160100400
