Learning Continuous Face Age Progression: A Pyramid of GANs.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biometrics

Background:

• Face age progression research often struggles to balance aging accuracy with identity permanence.
• Existing methods lack robust modeling of subject-specific characteristics and age-related facial changes.

Purpose of the Study:

• To present a novel generative adversarial network (GAN) approach for coupled face age progression.
• To ensure generated faces exhibit accurate aging effects while maintaining stable personalized properties.
• To advance the state-of-the-art in realistic and continuous face aging synthesis.

Main Methods:

• A novel GAN-based approach is proposed, separately modeling subject-specific traits and age-specific facial changes.
• A pyramidal adversarial discriminator estimates high-level age-specific features at multiple scales for realistic details.
• An adversarial learning scheme trains a single generator with multiple discriminators for smooth aging sequences.

Main Results:

• The proposed method achieves remarkably vivid aging effects, robust to variations in pose, expression, and makeup.
• Quantitative evaluations show accurate recovery of target age distributions.
• High verification rates (99.88% and 99.98%) were achieved on MORPH and CACD databases after significant age transformations.

Conclusions:

• The novel GAN approach effectively addresses the dual requirements of aging accuracy and identity permanence in face age progression.
• The method demonstrates superior performance in generating realistic and continuous age-progressed facial images.
• This work significantly advances the capabilities of face age progression technology.