Cross-GAN: Unsupervised Image-to-Image Translation
Xu Peng, Qishen Li, Taida Wu, Sihao Yuan · 2022 IEEE 6th Information Technology and Mechatronics Engineering Conference (ITOEC) · 2022
In recent years, unsupervised image-to-image translation is a topic in computer graphics. This paper proposes a new architecture to translate the image through the idea of sharing and decoupling, termed as Cross-GAN, which can better extract the style in the process of translation, and does not change the original outline and background. We improve the cycle consistency loss and reconstruction loss for the unique properties of the model to make the image more realistic. Because GAN[1] only generates the target style image by inputting the content image, and the information of the target domain image is not used in the training process. We input style images and content images through cross-input. The style images play a guiding role in the generator, and the effect of image are better by fusing style features and content features. Due to the fact that the image of the target domain is input, it is easy to produce the phenomenon of over-fitting. In order to prevent over-fitting training, we improve the cycle consistency loss and reuse the input encoder to make the model more compact.