Automated quantitative aesthetic evaluation of visual artworks is a challenging cross-disciplinary task involving computer vision and art history. Traditional aesthetic assessment methods rely on handcrafted features or single-branch deep learning models, which fail to comprehensively capture the multi-faceted artistic attributes (e.g., color harmony, composition balance, texture, and semantic style) and long-range global dependencies critical to artistic appreciation. To address these limitations, this paper proposes a novel framework: Vision Transformer with Multi-Dimensional Artistic Feature Fusion (MDAF-ViT). Our model integrates a hierarchical Vision Transformer (ViT) backbone for global context modeling with multi-branch feature extractors to capture low-level visual attributes, mid-level compositional rules, and high-level semantic style features. A key innovation is the Dynamic Multi-Dimensional Attention Fusion (MDAF) module, which adaptively weights and fuses heterogeneous artistic features. Extensive experiments on standard art aesthetic datasets (BAID, APDDv2, JenAesthetics) demonstrate that MDAF-ViT significantly outperforms state-of-the-art CNN and ViT-based methods, achieving superior performance in terms of Pearson Linear Correlation Coefficient (PLCC), Spearman Rank Correlation Coefficient (SRCC), and Mean Squared Error (MSE). This work provides a robust, interpretable foundation for large-scale digital art analysis and curation.
Zeyu Gao (Fri,) studied this question.