Chinese Painting Semantic Resource for Resolution-Aware Digitisation and Structured Heritage Access
TL;DR
The Chinese Painting Semantic Resource (CPSR), comprising 1,797 deduplicated painting-image records with a 12-column experiment-facing metadata release and a bilingual, 24-type lightweight semantic application profile, shows how local semantic distinctions can be retained while preparing for standards-aligned access.
Abstract
Chinese painting digitisation poses a distinctive heritage-informatics challenge: culturally salient visual structures such as brushwork, intentional blank space, and text–image coexistence are difficult to document, access, and research with general-purpose collection and computational representations. We present the Chinese Painting Semantic Resource (CPSR), comprising 1,797 deduplicated painting-image records with a 12-column experiment-facing metadata release and a bilingual, 24-type lightweight semantic application profile. Using single-seed classification experiments across six architectures as diagnostic probes of resized derivative inputs, we find that 512 \(\times\) 512 gives the best primary-category result (96.7%) and broad gains over 224 \(\times\) 224, whereas the effect of 1024 \(\times\) 1024 on seven merged technique–subject classes is model-dependent: it gives the best individual result (81.7%) but a 0.9 percentage-point lower cross-model mean than 512 \(\times\) 512. Three post-hoc Grad-CAM case studies provide exploratory, correlational observations rather than quantitative localisation validation. By linking culturally grounded profile design to representative CIDOC-CRM, Getty AAT, IIIF, and W3C Web Annotation mappings, CPSR shows how local semantic distinctions can be retained while preparing for standards-aligned access; it does not claim complete per-record profile population or model-level semantic understanding.