Style Transfer at 100+ FPS Via Sub-Pixel Super-Resolution
Haoyu Li, Xiangmin Xu, Bolun Cai, Kailing Guo, Xiaofen Xing · 2018
Recently, feed-forward style transfer networks have achieved comparable results with optimization-based methods. However, existing style transfer networks are still computationally inefficient in embedded device, and cannot handle well on fine or exquisite textures. Since a stylize image represents the semantic information of the content image with the style of the style image, it is unnecessary to use the whole content image for style transfer. In this paper, we propose a faster style transformation network (Faster-StyleNet), which takes a low-resolution content image as input, and generates the high-resolution stylized image by incorporating superresolution strategy. In the proposed Faster-StyleNet, high quality stylized images are produces at 100 + fps.