SiGAN: Siamese Generative Adversarial Network for Identity-Preserving Face Hallucination

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Generative adversarial networks (GANs) excel at hallucinating high-resolution (HR) faces from low-resolution (LR) inputs.
• A key limitation of existing GANs is the inability to preserve facial identity during hallucination, hindering recognition.
• This results in reconstructed HR faces that are visually plausible but lack identity consistency.

Purpose of the Study:

• To develop a novel GAN architecture that ensures identity preservation during face hallucination.
• To reconstruct high-resolution faces that are both visually realistic and recognizable.
• To improve the performance of face hallucination systems in terms of identity verification.

Main Methods:

• Proposed a Siamese GAN (SiGAN) architecture, featuring a pair of identical generators and a single discriminator.
• Incorporated reconstruction error and identity label information into the loss function in a pairwise manner.
• Iteratively optimized the loss functions for the generator pair and discriminator to achieve dual objectives.

Main Results:

• SiGAN successfully reconstructs HR faces that visually resemble the original identities.
• The method significantly outperforms existing face hallucination GANs in objective face verification.
• Achieved promising visual-quality reconstruction, even for unseen identities not present in the training data.

Conclusions:

• SiGAN effectively addresses the identity preservation challenge in GAN-based face hallucination.
• The proposed architecture balances visual quality with identity recognition accuracy.
• SiGAN demonstrates robust performance and potential for real-world applications requiring reliable face reconstruction.