Fine art image classification and design methods integrating lightweight deep learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Digital Art History

Background:

• Fine art image classification faces challenges with low efficiency and poor generalization.
• Existing deep learning models can be computationally intensive and struggle with subtle style variations.

Purpose of the Study:

• To develop an efficient and robust fine art image classification method using lightweight deep learning.
• To improve the accuracy and generalization capabilities of art image classification models.

Main Methods:

• A lightweight hybrid network, MobileNet-Transformer Hybrid (MTH), combining depthwise separable convolution and multi-head self-attention.
• A dynamic channel-spatial attention module (DCSAM) for adaptive feature enhancement.
• A cross-style feature transfer (CSFT) framework utilizing contrastive learning for improved robustness.

Main Results:

• The MTH model achieved high classification accuracy (85.2% on ArtBench-10) with significantly reduced parameters (1.2M).
• DCSAM effectively enhanced local style-discriminative features (brushstrokes, colors), reducing misclassifications of similar styles.
• CSFT improved generalization for rare styles in long-tailed datasets by constraining cross-style feature distances.

Conclusions:

• The proposed lightweight deep learning approach offers an efficient solution for fine art image classification.
• The method demonstrates practical value for art design automation and cultural heritage digitization.
• This study contributes theoretical innovation in efficient deep learning for specialized image analysis tasks.