Applying deep learning for style transfer in digital art: enhancing creative expression through neural networks

Area of Science:

• Computer Vision
• Digital Art
• Machine Learning

Background:

• Neural style transfer (NST) blends artistic styles with image content.
• Traditional NST methods struggle with balancing style fidelity, content preservation, and computational efficiency.
• Existing models often lack the speed required for real-time applications.

Purpose of the Study:

• To enhance the efficiency and quality of neural style transfer.
• To address challenges in content retention, style fidelity, and computational performance.
• To improve visual coherence and consistency for practical NST applications.

Main Methods:

• Proposed a refined model integrating Adaptive Instance Normalization (AdaIN) and Gram matrix-based style representation within a convolutional neural network (CNN).
• Evaluated the model using quantitative metrics: content loss, style loss, Structural Similarity Index (SSIM), and processing time.
• Conducted qualitative assessments of content and style consistency.

Main Results:

• Achieved a 15% reduction in content and style loss compared to baseline models.
• Obtained an SSIM score of 0.88 for medium style intensity, indicating maintained structural integrity.
• Reduced processing time by 76%, enabling near-real-time performance.
• Maintained high style fidelity across various artistic styles with minimal content loss.

Conclusions:

• The refined NST model effectively balances style and content, enhancing visual quality and computational efficiency.
• Advancements make NST more accessible for real-time artistic applications.
• The model serves as a versatile tool for digital art, design, and multimedia production.