A Foreground-Aware Framework for Local Face Attribute Transfer.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Digital Image Processing

Background:

• Headshot photography is prevalent on social media, but professional editing is complex for average users.
• Existing style transfer algorithms lack semantic understanding and local control over facial attributes.
• There's a need for methods that allow targeted facial feature modification while maintaining image integrity.

Purpose of the Study:

• To develop a novel framework for semantically meaningful local face attribute transfer.
• To enable flexible transfer of specific facial organ attributes from a reference to an input image.
• To preserve the original background during attribute transfer for photorealistic results.

Main Methods:

• A novel framework for local face attribute transfer was developed.
• Reference images are warped to match the input image's shape, pose, location, and expression.
• A neural style transfer algorithm is initialized using a fusion of warped and input images.

Main Results:

• The proposed method achieved superior performance with an inception score of 3.81 and Fréchet inception distance of 80.31.
• Results indicate a 10% improvement over competing methods in producing high-quality, photorealistic attribute transfer.
• Experimental and theoretical findings demonstrate the framework's efficacy and superiority.

Conclusions:

• The developed framework offers effective and semantically meaningful local face attribute transfer.
• It provides a flexible solution for users to modify specific facial features while preserving backgrounds.
• The method outperforms state-of-the-art alternatives in generating realistic and high-quality results.