WeaGAN:Generative Adversarial Network for Weather Translation of Image among Multi-domain

Abstract

Weather translation of image refers to the task of changing the weather of an input image to desired weather while preserving the structure of the image's content, which belongs to a task of image-to-image translation. Recent works have made great process in image-to-image translation between two domains and some works have even achieved multi-domain translation within a single model. However, existing works have limited robustness in handling weather translation among multi-domain, since bad weather produces a loud noise and it is challenging to process scene images without fixed pattern in a unified model. In this paper, we propose WeaGAN based on encoder-decoder architecture and generative adversarial training process to translate the weather of image among multi-domain. In particular, We employ SE block in generator and combine adversarial loss, classification loss and content loss for visually detailed and realistic result. Experience in qualitative and quantitative aspect on synthetic dataset and real dataset show the effectiveness and competitiveness of our method compared with state-of-the-art works.