Fine-Grained Multi-Attribute Adversarial Learning for Face Generation of Age, Gender and Ethnicity

Abstract

Since the Generative Adversarial Network (GAN) was proposed, facial image generation used for face recognition has been studied in recent two years. However, there are few GAN-based methods applied for fine-grained facial attribute analysis, such as face generation with precise age. In this paper, fine-grained multi-attribute GAN (FM-GAN) is presented, which can generate fine-grained face image under specific multiply attributes, such as 30-year-old white man. It shows that the proposed FM-GAN with fine-grained multi-label conditions is better than conditional GAN (cGAN) in terms of image visual fidelity. Besides, synthetic images generated by FM-GAN are used for data augmentation for face attribute analysis. Experiments also demonstrate that synthetic images can assist the CNN training and relieve the problem of insufficient data.