Edge-guided Adversarial Network Based on Contrastive Learning for Image-to-Image Translation

In recent years, generative adversarial networks have made great progress in image synthesis and image translation tasks in the field of image processing and computer vision. However, the quality of the generated image and the scalability over multiple datasets is still not satisfying. We briefly review some prior works and propose a method for image-to-image translation, which is learning a mapping between different visual domains. The network extracts edge feature from both domains of output and target, and minimizes the difference using a framework based on patchwise contrastive learning. We apply edge feature guidance in our method and select Sobel operator among several classical edge detection operators. We demonstrate that our method outperforms existing approaches in the task of unpaired image-to-image translation across datasets.