Computer Vision and Image Analysis of Art 2026

MONDAY 2 MARCH 2026

Computer Vision and Image Analysis of Art

08:30 - 10:30
Grand Peninsula G

08:30CVAA-172
Leveraging vision language models for semantic interpretation of historical paintings: A case study on religiousness, Yuya Kanazawa, Chiba University (Japan); Midori Tanaka, Chiba University (Japan); Hiroshi Kera, Chia University (Japan); Takahiko Horiuchi, Chiba University (Japan) [view abstract]

This study investigates the use of Vision-Language Models (VLMs) to automatically extract the religious contexts embedded within historical paintings and to quantitatively evaluate the reliability of such interpretations. While prior research has focused largely on surface-level features such as style classification or conservation techniques, few methodologies have addressed abstract and symbolic meanings such as religiosity. In this work, curator-authored explanatory texts were obtained via the Cleveland Museum of Art's public API. These texts were then compared against outputs generated by four VLMs--LLaVA-1.5, LLaVA-NeXT, Qwen-VL-Chat, and GPT-4V--when given prompts such as "Explain the religious significance of this painting." The similarity between the model outputs and curator texts was assessed using BERTScore, ROUGE, and BLEU, providing a multifaceted evaluation of semantic alignment. Results indicate that while VLM outputs tend to lack comprehensiveness, they exhibit strong precision, selectively capturing plausible religious elements. By establishing a methodological framework for evaluating religiosity in art, this study lays the groundwork for extending VLM-based interpretation to broader contextual elements such as emotional expression, cultural symbolism, and socio-political messages, thereby fostering deeper integration between engineering approaches and the digital humanities.

08:50CVAA-173
A forward-looking multi-factor authentication -MFA- model for cultural heritage art objects that can be trained to look backwards, Larry Kleiman, Spectral Masters Digital Imaging, Inc. (US Territories and Minor Outlying Islands) [view abstract]

This paper proposes a scientifically rigorous Multi-Factor Authentication (MFA) framework for cultural heritage art objects, introducing a standardized methodology for establishing authenticity based on inherence factors unique physical properties intrinsic to the object itself. The proposed MFA model integrates three primary inherence factors Visual Perception Color Matching (VPCM)1, Z-Score statistical surface topography analysis, and Shifted-Excitation Raman Difference Spectroscopy (SERDS) to form a Basic Authentication Panel. A complementary Legacy Panel employing X-Ray Fluorescence (XRF), Ultraviolet Fluorescence (UVF), and Reflective Fourier Transform Infrared (FTIR) analysis extends the model to historic artworks. Together, these methods create a reproducible, data-driven authentication system that can train artificial intelligence (AI) models to both authenticate future works and retrospectively assess legacy art. The result is a forward-looking protocol grounded in traceable, non-destructive scientific measurement and secured through blockchain-based data archiving.