Oil-Painting Style Classification Using ResNet with Conditional Information Bottleneck Regularization.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Art History

Background:

• Automatic classification of oil painting styles is crucial for art history, digital archiving, and forensic analysis.
• Existing methods often lack the fine-grained discrimination needed for nuanced artistic style identification.

Purpose of the Study:

• To introduce a novel deep conditional information bottleneck (CIB) framework for accurate and interpretable fine-grained oil painting style classification.
• To enhance the analysis of visual artistic attributes in oil paintings.

Main Methods:

• Developed a deep conditional information bottleneck (CIB) framework utilizing ResNet-50 architecture.
• Minimized conditional mutual information I(X;Z∣Y) using a matrix-based Rényi's entropy estimator for computational efficiency.
• Evaluated the CIB framework on the Pandora and OilPainting datasets.

Main Results:

• The CIB framework achieved superior performance compared to standard ResNet on both datasets.
• Demonstrated a relative performance gain of 13.1% on the Pandora dataset and 11.9% on the OilPainting dataset.
• Generated more disentangled latent representations that effectively cluster semantically similar artistic styles.

Conclusions:

• The proposed CIB framework offers a computationally efficient and effective solution for fine-grained oil painting style classification.
• The method provides enhanced interpretability through semantically clustered latent representations.
• This approach advances objective and scalable analysis of visual artistic attributes.