GAN-based facial attribute editing and its applications to face recognition

Abstract

[en] Face recognition has achieved impressive results in recent years using Deep Learning models. Deep face recognition has seen fast-paced progress since 2014, the year in which the first deep face recognition method was proposed (Deepface [1]). These models have become more robust to unconstrained situations, becoming suitable for in-the-wild conditions. This robustness has been demonstrated by achieving very good results in popular face recognition benchmarks. Nevertheless, there is still room for improvement when it comes to the way they are trained. Particularly, datasets that are created to train face recognition models are biased and are not as diverse as needed. Most datasets are created from celebrities on formal occasions, hence people who appear in these images are majorly young people smiling and wear make-up. To address this problem, we propose to use a data augmentation technique based on Generative Adversarial Networks (GANs), which are a type of generative model. GANs [2] have been proven successful in many computer vision tasks such as photo-realistic image generation [3, 4], single and multidomain image-to-image translation [5, 6, 7], and super resolution imaging [8]. They have also been applied to facial attribute editing (FAE), which consists of changing people’s facial attributes (e.g., synthesize bangs, add sunglasses, change one’s age). The main research objective of this work will be to analyze the effect of these GANs for FAE in the training of face recognition models. To do so, three phases will be carried out in this project. In the first phase, we will explore the state-of-the-art GANs for FAE and retrain them with a better-suited dataset for our data augmentation technique. We will emphasize the importance of changing the training dataset for our purposes. In the second phase, we will retrain these GANs with two more losses to improve the quality of these generative models. We will show how these two losses help, analyzing them quantitatively and qualitatively. In the last phase, we will apply the GAN-based data augmentation technique with the models that were trained in previous phases. To analyze the effect of this technique, we will design a set of experiments with the objective of highlighting the benefits of our proposed data augmentation on face recognition models. To conclude, we will briefly comment on the advantages and disadvantages of this technique and possible lines of future work.