MSGAN: Generative Adversarial Networks for Image Seasonal Style Transfer
Fuquan Zhang, Chuansheng Wang
Abstract
Although Generative Adversarial Networks (GANs) have shown remarkable successes in various computer vision tasks, they still face challenges in image season style transfer task. In this paper, we propose a multi-season Generative Adversarial Networks (MSGANs) aimed to transfer input images into other season styles. To improve the quality of the simulated images generated by the proposed MSGAN, we propose a novel loss function to guide the optimization direction of the network. Besides, we adopt the saliency information to guide the seasonal style transformation task, so as to ensure that different image contents can have different optimization weights in MSGAN. The experimental results show that the proposed MSGAN can generate high-quality simulated images from real images, and is superior to other latest methods. Not only that, the synthetic image generated by the proposed method also be used to perform depth estimation task so that prove that the synthetic images can be well applied to other computer vision tasks.