FNR-GAN: Face Normalization and Recognition with Generative Adversarial Networks
Abstract
The normalization of faces in the wild is an interesting process which can improve face recognition performances and avoid the complex computation of face generation in cross different variations. In this paper, we propose a multi-objective approach that generates and recognizes normalized faces while preserving their identities. An unsupervised Face Normalization and Recognition framework using discriminant normalized features is presented. This latter is based on an optimized combination of Generative Adversarial Network (GAN) generators and Convolutional Neural Network (CNN) classifiers. The main power of our approach is to generate optimized features representing normalized faces finding a trade-off between improving identity preservation and minimizing the architecture complexity. Additionally, it can be adapted to impaired and unlabeled datasets which can respond to real-world face variations and available data. Experimental results show that the proposed method outperforms other models on face normalization and achieves state-of-the-art frontal-frontal face verification in CFP protocol and face recognition in LFW. The code and results are available at github/FNR-GAN.