Unpaired Artistic Portrait Style Transfer via Asymmetric Double-Stream GAN

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Portrait style transfer is a growing research area within image style transfer technologies.
• Existing methods like cartoonization often result in undesirable artifacts such as facial deformation and loss of contours in portrait images.

Purpose of the Study:

• To address the limitations of current portrait style transfer techniques, including facial deformation and contour issues.
• To develop a novel generative adversarial network for unpaired artistic portrait style transfer that preserves crucial facial features.

Main Methods:

• An asymmetric double-stream generative adversarial network (ADS-GAN) was developed.
• An edge contour retention (ECR) regularized loss was proposed to maintain local and global contours, preventing deformation.
• A content-style feature fusion module with a style attention mechanism was integrated to enhance style learning.
• A guided filter was incorporated into the content encoder to preprocess source images, reducing negative impacts on style transfer.

Main Results:

• The proposed ADS-GAN method demonstrated superior performance compared to benchmark methods in qualitative and quantitative analyses.
• The generated portraits successfully preserved the overall structure and contours of the original images.
• Ablation and parameter studies confirmed the effectiveness of individual components within the ADS-GAN framework.

Conclusions:

• The developed ADS-GAN is effective for unpaired artistic portrait style transfer.
• The integration of ECR loss and style attention mechanism significantly improves the quality and fidelity of style-transferred portraits.
• The method successfully mitigates common issues like deformation and contour loss in portrait style transfer.