Generative Data Augmentation applied to Face Recognition

In this paper, we present a data augmentation method whose goal is to generate face images and maximize faces variation in the training set. The main objective is to break free from the traditional data augmentation techniques used in deep neural networks such as geometric and photometric transformations. Our method consists in generating face images using Deep Convolutional Generative Adversarial Networks (DC-GAN) feed with light pose variations of the face in 2D plane. Its a selective feature space augmentation. Then, we apply face resolution enhancement based on Enhanced Super Resolution GAN (ESRGAN), since the generated faces are inferior and noisy. As a final step, we perform face verification using Deep Convolutional Neural Networks (CNNs) to confirm the robustness of the used pipeline. The found results achieves comparable performance in comparison with the state-of-the-art methods.