Arbitrary Style Transfer via Learning to Paint in the Feature Domain
Yujie Huang, Yixuan Liu, Minge Jing, Mingyu Wang, Xiaoyong Xue, Xiaoyang Zeng, Yibo Fan · 2021 IEEE 14th International Conference on ASIC (ASICON) · 2021
Style transfer is enjoyed by people in multimedia fields, like photography. However, arbitrary style transfer still faces the challenge of fully migrating the style while maintaining content consistency. To solve this, we propose a novel arbitrary style transfer algorithm named LPFD. We firstly propose the point of view of Learning to Paint in the Feature Domain (VLPFD). Then, according to the VLPFD, the Parameter Generation Network and Painting Network is designed to stylize the content feature map. Thanks to the VLPFD, the parameters of the Painting Network are significantly reduced, which considerably alleviates the burden of the Parameter Generation Network. The experimental results demonstrate that our algorithm can achieve more satisfactory results than the state-of-the-art arbitrary style transfer algorithms. Besides, LPFD can run in real-time.