Hierarchical Causal Learning for Face Age Synthesis.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Face age synthesis (FAS) models often struggle with feature disentanglement, leading to inaccuracies.
• Current methods entangle age-related features, hindering causal reasoning in facial image generation.

Purpose of the Study:

• To propose a hierarchical causal learning model (HCFace) for improved face age synthesis.
• To enhance feature disentanglement and causal reasoning in facial attribute manipulation.

Main Methods:

• Developed a hierarchical causal learning model (HCFace) integrating hierarchical structures and causal relationships.
• Designed a novel nonlinear mapping function to capture age-related facial attribute changes accurately.
• Leveraged hierarchical causal relationships for feature disentanglement in generative models.

Main Results:

• HCFace demonstrated superior performance compared to advanced baseline methods.
• Achieved an overall accuracy improvement of 2.47% in face age synthesis.
• Showcased significant improvements in age-related attributes like skin (9.75%) and hair (9.69%).

Conclusions:

• The proposed HCFace model effectively disentangles facial features for accurate age synthesis.
• Hierarchical causal learning enhances the causal reasoning capabilities of facial generative models.
• HCFace offers a promising approach for realistic face age manipulation and prediction.