DisP+V: A Unified Framework for Disentangling Prototype and Variation From Single Sample per Person.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics

Background:

• Single sample per person face recognition (SSPP FR) faces challenges due to limited enrollment data.
• Existing prototype plus variation (P+V) models struggle with nonlinear variations and noisy data.
• A need exists to disentangle identity (prototype) from variations in a latent space for robust face recognition.

Purpose of the Study:

• To propose a novel Disentangled Prototype plus Variation (DisP+V) model for SSPP FR.
• To address limitations of classic P+V models in handling nonlinear variations and data contamination.
• To enable semantic manipulation of face images in a disentangled latent space.

Main Methods:

• Developed an encoder-decoder generator and two discriminators for adversarial training.
• Implemented nonlinear encoding into a latent semantic space.
• Disentangled discriminative prototype features from less discriminative variation features.
• Utilized disentangled features to guide generation of identity-preserved prototypes and variations.

Main Results:

• The DisP+V model demonstrated superior performance over the classic P+V model on real-world face datasets.
• Achieved robust SSPP FR even with limited or noisy enrollment data.
• Showcased effectiveness in prototype recovery and face editing/interpolation tasks.

Conclusions:

• The proposed DisP+V model offers a significant advancement in SSPP FR.
• Disentangling features in a latent space enhances robustness to variations and data quality issues.
• DisP+V provides a versatile framework for face recognition and semantic image manipulation.