AD-VAE: Adversarial Disentangling Variational Autoencoder.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics

Background:

• Face recognition (FR) is crucial for security and access control but faces challenges with single-image datasets and variations like pose, illumination, and occlusion.
• Deep learning, including Variational Autoencoders (VAE) and Generative Adversarial Networks (GAN), has shown promise in FR.
• Single Sample Per Person Face Recognition (SSPP FR) specifically struggles with learning robust, identity-preserving features.

Purpose of the Study:

• To propose a novel framework, AD-VAE, to address the challenges in SSPP FR.
• To develop a method capable of learning representative, identity-preserving prototypes from diverse datasets.
• To effectively handle variations such as pose, illumination, and occlusion in face recognition.

Main Methods:

• The AD-VAE framework combines Variational Autoencoder (VAE) and Generative Adversarial Network (GAN) techniques.
• It employs four networks: an encoder and decoder (VAE-like), a generator for prototype creation, and a multi-task discriminator.
• The framework learns to build identity-preserving prototypes from controlled and uncontrolled datasets.

Main Results:

• AD-VAE significantly outperforms existing state-of-the-art face recognition techniques.
• Achieved high recognition rates on controlled datasets: AR (84.9%), E-YaleB (94.6%), CAS-PEAL (94.5%), and FERET (96.0%).
• Demonstrated remarkable performance on the uncontrolled Labeled Faces in the Wild (LFW) dataset with a 99.6% recognition rate.

Conclusions:

• The AD-VAE framework offers a robust solution for SSPP FR, effectively managing variations.
• It demonstrates superior performance compared to current methods on both controlled and uncontrolled datasets.
• AD-VAE holds significant potential for advancing face recognition research and real-world applications.