Edge Sensitive Unsupervised Image-to-Image Translation

The goal of unsupervised image-to-image translation (IIT) is to learn a mapping from the source domain to the target domain without using paired image sets. Most of the current IIT methods apply adversarial training to match the distribution of the translated images to the distribution of the target images. However, this may create artifacts in uniform areas of the source image if two domains have different background distribution. In this work, we propose an unsupervised IIT method that preserves the uniform background information of the source images. The edge information which is calculated by Sobel operator is utilized for reducing the artifacts. The edge-preserving loss function, namely Sobel loss is introduced to achieve this goal which is defined as the L2 norm between the Sobel responses of the original and the translated images. The proposed method is validated on the jellyfish-to-Haeckel dataset. The dataset is prepared to demonstrate the mentioned problem which contains images with different uniform background distributions. Our method obtained a clear performance gain compared to the baseline method, showing the effectiveness of the Sobel loss.