Enhancing artistic style classification through a novel ArtFusionNet framework.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Art History

Background:

• Artistic style classification is crucial for cultural heritage preservation and art retrieval.
• Convolutional Neural Networks (CNNs) excel at local texture analysis but struggle with long-range dependencies.
• Transformer models capture global context but lack fine-grained feature extraction.

Purpose of the Study:

• To develop a novel hybrid deep learning model, ArtFusionNet, for accurate artistic style classification.
• To synergize the strengths of CNNs and Transformers for improved feature representation.
• To achieve state-of-the-art performance in classifying diverse artistic styles.

Main Methods:

• Developed ArtFusionNet, integrating an Adaptive Fusion Module (AFM) to fuse CNN and Transformer features.
• Employed dilated convolutions and pyramid pooling for hierarchical CNN feature extraction.
• Utilized tokenization and multi-head self-attention for Transformer-based global modeling.
• Implemented learnable weighting within the AFM for optimal feature fusion.

Main Results:

• Achieved a state-of-the-art accuracy of 99.00% on multiple datasets (Fallah_artist_dataset, WikiArt, BAM!, Painting-91).
• Outperformed stand-alone CNNs, Transformers, and previous hybrid architectures.
• Ablation studies confirmed mutual reinforcement between CNN and Transformer components.
• Statistical tests (t-test, ANOVA) confirmed model robustness and balanced performance.

Conclusions:

• ArtFusionNet effectively links local and global feature modeling for advanced artistic style classification.
• The hybrid approach offers significant advancements in computer vision for art analysis.
• Future work includes model compression, self-supervised learning, and multi-modal integration for enhanced efficiency and generalization.