Hierarchical Multi-Stream Feature Network for Digital Art Aesthetic Quality and Style Classification Through Intelligent Systems

Abstract

Digital art analysis is evolving rapidly, with intelligent systems playing a growing role in understanding aesthetic quality and artistic styles. In this work, we present the Hierarchical Multi-Stream Feature Network (HMSFN), a deep learning framework designed to improve the way visual features are extracted and classified across different styles and aesthetic levels. The study is based on a curated dataset of 213,000 digital artworks sourced from online galleries and collections, covering a wide range of creative expressions and thematic categories. To enhance data quality and balance, we applied specialized preprocessing techniques including Contrast-Balanced Normalization, Dominant Color Mapping, and Gradient-Symmetric Scaling. Additionally, Weighted Synthetic Feature Augmentation (WSFA) was introduced to address class imbalance, while an Adaptive Feature Filtering Framework (AFFF) was used to remove redundant features and retain the most informative ones. The model was trained using an 80:20 split and evaluated against several leading deep learning approaches. HMSFN, which combines DenseNet, ConvNeXt, and Vision Transformer in a multi-stream configuration, achieved outstanding results—99.0% accuracy, 98.6% F1-score, 97.5% LCCR, and an AUC of 99.3%. These findings highlight the effectiveness of our approach in capturing complex visual attributes and support its use in digital art classification and computational aesthetics.