Generating photo-realistic training data to improve face recognition accuracy

Face recognition has become a widely adopted biometric in forensics, security and law enforcement thanks to the high accuracy achieved by systems based on convolutional neural networks (CNNs). However, achieve good performance, CNNs need be trained with very large datasets which are not always available. In this paper we investigate feasibility of using synthetic data augment face datasets. particular, propose novel generative adversarial network (GAN) that can disentangle identity-related attributes from non-identity-related attributes. This is done training an embedding maps discrete identity labels latent space follows simple prior distribution, GAN conditioned samples distribution. A main novelty our approach ability generate both images subjects set new set, use By recent advances training, show generated model photo-realistic, augmented those lead increased accuracy. Experimental results method more effective when augmenting small absolute improvement 8.42% was dataset less than 60k facial images. • Generate photo-realistic conditional GAN. Two vectors encode non-related respectively. Map features continuous space. Training sets better balance between real outperformed.