Semi-supervised Adversarial Image-to-image Translation

摘要

Image-to-image translation involves translating images in one domain into images in another domain, while keeping some aspects of the image consistent across the domains. Image translation models that keep the category of the image consistent can be useful for applications like domain adaptation. Generative models like variational autoencoders have the ability to extract latent factors of generation from an image. Based on generative models like variational autoencoders and generative adversarial networks, we develop a semi-supervised image-to-image translation procedure. We apply this procedure to perform image translation and domain adaptation for complex digit datasets.