A hierarchically trained generative network for robust facial symmetrization.

Area of Science:

• Computer Vision
• Medical Imaging
• Artificial Intelligence

Background:

• Facial symmetry is crucial in medical and academic fields, including diagnosing facial disorders and evaluating beauty.
• Existing automatic facial symmetrization methods often require manual intervention and are sensitive to image conditions.
• Limited research exists for fully automatic facial symmetrization, hindering its broader application.

Purpose of the Study:

• To develop an automatic facial symmetrization method that is realistic and preserves facial identity.
• To overcome limitations of existing methods, such as manual intervention and sensitivity to image quality.
• To present a novel deep generative network model for enhanced facial symmetrization.

Main Methods:

• A joint-loss enhanced deep generative network model is proposed for full facial image analysis.
• The joint-loss incorporates adversarial losses for realism and an identity loss for identity preservation.
• A multi-stage training process is employed, avoiding the need for large datasets of symmetrical faces.

Main Results:

• The proposed model achieves competitive results compared to existing methods based on facial landmark detection.
• The method demonstrates effective facial symmetrization while maintaining the original identity.
• The multi-stage training approach successfully avoids the requirement for extensive symmetrical face training data.

Conclusions:

• The joint-loss enhanced deep generative network offers a robust solution for automatic facial symmetrization.
• This approach improves realism and identity preservation in symmetrized facial images.
• The method shows promise for applications in medical diagnosis and facial analysis.