Haar Wavelet-Based Representation Learning for Unpaired Image-to-Image Translation

In recent years, there have been numerous attempts to achieve unpaired image-to-image translation. Many algorithms have especially incorporated the contrastive learning framework into unpaired image-to-image translation. This paper presents an innovative approach to applying contrastive learning to unpaired image-to-image translation using the Haar wavelet transform, called HAWAII. Specifically, we investigated how mutual and domain-specific information can be optimally chosen for contrastive learning. This leads to a novel definition of typical and non-typical features by using features from specific frequency bands that are separated using the Haar wavelet transform. By utilizing the typical and non-typical features for contrastive learning, the generator can effectively account for which information should be preserved and which should be modified for image translation. Since contrastive learning is effective for the classification task, we applied it not only to the generator but also to the discriminator. The proposed regularization method enables the model to efficiently identify the information that should be retained in the given image and the translated image in the generator and to better distinguish the features of the source and target domains in the discriminator. Extensive experiments on various tasks demonstrate the feasibility of the proposed method, which achieves state-of-the-art performance in terms of many metrics.