动漫插画风格图像生成研究

Image-to-Anime Translation via Generative Adversarial Networks

把人像或景物等真实世界中的图片直接变换为动漫或插画风格图像是一件非常有趣且具有实际应用需求的任务,自该概念提出后就在计算机视觉领域引起了广泛的研究兴趣。当前在该任务的研究中主要通过生成对抗学习模型对图像进行风格迁移来实现。但是现有方法在图像风格迁移时难以同时传输风格与内容。本文针对该问题设计了新的生成器网络,在同时迁移风格与内容时做到了比较好的平衡。在将设计的生成对抗模型应用到图像风格转换任务上时,能够根据目标艺术风格生成对应的动漫插画效果。这些生成的插画图像不仅拥有非常好的视觉质量,而且原有的图像内容也能被很好地保留。

It is an interesting task to convert real world images, such as portraits and scenery, into corresponding anime or illustrations. This job, which is also named image style transfer, has attracted great attentions in computer vision community. Current generative adversarial learning based deep models fail to transfer style and content at the same time. This paper proposes a new generator network to get a better balance when transferring style and content together. The proposed new generative adversarial model is also applied to the image style transfer task. Experimental results show that it can generate excellent illustrations according to the target artistic style. The resulting illustrations have higher visual quality and preserve the original information well.