Single Sample Face Recognition Based on Identity-Attribute Disentanglement and Adversarial Feature Augmentation

Abstract

To address the issue of facial variation interference, this paper proposes a novel approach for single sample face recognition. Inspired by human visual perception, we introduce an attribute disentanglement module to separate identity features from attribute features using canonical correlation analysis. Due to the lack of attribute labels in the single sample set, we utilize the attribute features of the generic set to construct the SOM attribute space. Then, we fine-tune the network by reducing the distance between the attribute features of single sample and the attribute space. Finally, we use feature adversarial augmentation module to generate more intra-class features and train more robust classifier. Experimental results on AR, LFW and FERET datasets show significant improvements in accuracy and generalization performance compared to other methods.

姚亮

Master of Engineering

My research interests include Artificial Intelligence, UAV Object Detection and Feature Disentanglement.