Soft-biometrics encoding conditional GAN for synthesis of NIR periocular images

Abstract

Soft-Biometric information, such as gender, has great potential for applications in security, forensics and marketing. Unfortunately, there are few gender-labelled databases available which make state of the art techniques, such as deep learning, difficult to use. An alternative source of data to train these algorithms are synthetic images. Methods based on Generative Adversarial Network are widely used for generating synthetic images. However, low features, such as gender, are not preserved in the images generated by these methods. In this paper, a novel GAN-based algorithm that preserves gender information while generating synthetic images is presented. It uses a latent vector that encodes gender information within the conditional GAN algorithm. Resulting synthetic images were tested using a gender classifier algorithm (CNN). Experiments demonstrate that the proposed method can be a useful tool for the synthesis of gender-labelled images to be used in training Deep Learning gender-classification algorithms. As an additional contribution a novel person-disjoint gender labelled dataset is presented (UNAB-Gender).