Computational aesthetics for digital interpretation of traditional art and cultural heritage: a two-level benchmark using AVA and heritage image datasets

Computational aesthetics is a promising approach to connect deep visual learning with the digital interpretation of traditional art and cultural heritage. Existing image aesthetic assessment research has primarily concentrated on general photographic databases, and the transferability of image aesthetic representations learned from those databases to culturally specific heritage images has not been well studied. In this study, the proposed two-level computational aesthetics framework includes learning general aesthetic representations from the Aesthetic Visual Analysis dataset and transferring them to four heritage image datasets: Cultural Heritage Visual Dataset, Indian Paintings Dataset, and Batik Nusantara Dataset, . ResNet50, EfficientNet B3, ViT B/16, Swin T, and CLIP ViT B/32 are assessed on multi-task aesthetic learning tasks, such as high- and low-aesthetic classification, mean-score regression, and vote-distribution prediction, at Level 1. The learned aesthetic features are adapted to the heritage domains in Level 2 by freezing the features extracted from the image within the network, performing full fine-tuning, and adding an aesthetic-aware fusion module. The fusion module projects and normalizes the visual and aesthetic embeddings using softmax and layer normalization, and learns the projection weights. Additionally, a CLIP semantic branch and an explainability module using Grad CAM, attention rollout, and semantic similarity maps are integrated to support interpretable cultural analysis. The proposed framework is a reproducible standard for evaluating the potential of general aesthetic knowledge to enhance computational interpretations of traditional art and heritage images across domains.