Age-Oriented Face Synthesis With Conditional Discriminator Pool and Adversarial Triplet Loss.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Synthesis

Background:

• Generative Adversarial Networks (GANs) are widely used for realistic face image generation.
• Vanilla GANs struggle with mode collapse, leading to poor synthesis quality and limited visual variation.
• Existing methods for age-oriented synthesis have limitations in preserving identity information.

Purpose of the Study:

• To propose a novel method for age-oriented face synthesis.
• To achieve high synthesis accuracy and strong identity permanence capabilities.
• To address the mode collapse issue and enhance identity preservation in synthesized faces.

Main Methods:

• A Conditional Discriminator Pool with age-specific discriminators to combat mode collapse.
• A novel Adversarial Triplet loss, enhancing the Triplet loss with a ranking operation for reduced intra-class variance.
• Utilizing these components for accurate and identity-preserving age-oriented face synthesis.

Main Results:

• The proposed method significantly outperforms state-of-the-art methods in age-oriented face synthesis.
• Demonstrated improvements in both synthesis accuracy and identity permanence capabilities.
• Qualitative and quantitative experimental results validate the effectiveness of the approach.

Conclusions:

• The novel method effectively tackles mode collapse and enhances identity permanence in face synthesis.
• Achieves superior performance compared to existing techniques in age-oriented face generation.
• Presents a promising advancement for realistic and identity-preserving facial image synthesis.