Information-Cycle Dual Contractive Learning with Regularization for Unsupervised Image-to-Image Translation

Abstract

In the tasks of image-to-image translation, we aim to add the appearance features of the target domain on the basis of retaining the structural features of the input image. Contrastive learning methods are gradually applied to image translation tasks. Despite the previous progress in image translation models, it remains challenging to employ an efficient learning setting by capturing image features from multiple perspectives to build mappings. In this paper, we propose a novel method based on information-cycle dual contrastive learning with regularization (ICDCLGAN-R) to improve the overall quality and visual observability which mainly consists of three components: dual-domain GAN, contrastive regularization (CR) and information cycle. Specifically, dual-domain GAN contributes to learning cross-domain image mappings and capturing the domain gap more efficiently by maximizing mutual information between corresponding patches of input and output. Simultaneously, CR utilizes two images from one domain (one real image and one generated image) and one generated image belonging to another domain as positive, anchor and negative, respectively, aiming to pull positive pairs in some metric space and push apart the representation between negative pairs. CR compares multi-source image features meanwhile enhancing information reusability. Besides, we put generated images into dual-domain GAN ulteriorly to perform additional iteration. It reinforces and verifies reconfigurability between images from different domains. In our experiments, we demonstrate that our method improves the quality of generated images and utilizes image features more efficiently.