Conditional Perceptual Adversarial Variational Autoencoder for Age Progression and Regression on Child Face

Abstract

Recent works have shown that Generative Adversarial Networks (GAN) and Variational Auto-Encoder (VAE) can construct synthetic images of remarkable visual fidelity. In this paper, we propose a novel architecture based on GAN and VAE with Perceptual loss termed as Conditional Perceptual Adversarial Variational Autoencoder (CPAVAE), a model for face aging and rejuvenation on children face. CPAVAE performs face aging and rejuvenation by learning manifold constrained with conditions such as age and gender, which allows it to preserve face identity. CPAVAE uses six networks; these networks are an Encoder (E) and Sampling (S) which maps the child face to latent vector, Generator (G) takes the latent vector z as input along with age conditioned vector and tries to reconstruct the input image, a perceptual loss network Φ, a pre-trained very deep convolution network, discriminator on the encoder (Dz) smoothen’s the age transformation, discriminator on the image (Dimg) forces the generator to produce human realistic images. Here D and E are based on Variational Auto-encoder (VAE) architecture, VGGNet is used as perceptual loss network (Ploss), Dz and Dimg are convolutional neural networks. We represent child face progression and regression on the Children Longitudinal Face(CLF) dataset containing 10752 faces images in the age group [0 : 20]. This dataset contains 6164 and 4588 images of boys and girls respectively.