Unsupervised Cross Domain Learning for Noise Removal from a Single Image

Abstract

Most supervised deep learning-based frameworks designed for image noise removal usually exhibit robust denoising performance depending on large scale clean/noisy paired training images. However, paired training data may be unavailable in several real scenarios. In recent, cross domain transferring has been applied to unsupervised learning for image restoration. To avoid possible domain shift problem in existing cross domain learning frameworks, in this paper, an unsupervised cross domain deep learning framework is proposed for noise removal from a single image. Our goal is to learn an image generator without paired training data to directly learn invariant feature representations from noisy images and generate the corresponding clean images. In our framework, we aim at learning two image generators to transfer noisy images to clean images as well as clean images to noisy images, respectively, denoted by noise-to-clean and clean-to-noise generators, based on unpaired training images. To train this cross domain learning model, we propose a generative adversarial network (GAN)-based network architecture with different types of discriminators and loss functions. As a result, both generators can be efficiently trained, and the learned noise-to-clean generator can robustly and effectively perform feature learning from input noisy images and produce the corresponding denoised images.