Geometrically Editable Face Image Translation With Adversarial Networks.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Image-to-image translation methods often overlook geometric information, limiting their ability to edit shapes and structures.
• Existing models struggle with low-fidelity synthesis, particularly for complex subjects like human faces.

Purpose of the Study:

• To develop a novel image-to-image translation framework capable of geometric content editing.
• To enhance the realism and fidelity of synthesized images, especially for facial images.

Main Methods:

• Proposed Geometrically Editable Generative Adversarial Networks (GEGAN) for multi-domain mappings in geometric and attribute directions.
• Leveraged facial semantic segmentation to guide geometric editing within a variational autoencoder framework.
• Employed multi-scale regional discriminators to focus on critical facial components for improved detail.

Main Results:

• Demonstrated successful geometric modification capabilities in face image translation.
• Achieved significant improvements in the fidelity and realism of synthesized images compared to existing methods.
• Validated performance through quantitative and qualitative evaluations on the CelebA dataset.

Conclusions:

• GEGAN effectively addresses the limitations of current image-to-image translation techniques by incorporating geometric control.
• The proposed method offers a powerful tool for realistic and editable face image synthesis.
• Facial semantic segmentation is a key component for achieving high-fidelity geometric manipulation in generative models.