Abstract
Modern research has demonstrated that many eye-catching images can be generated by style transfer via deep neural network. There is, however, a dearth of research on content-aware style transfer. In this paper, we generalize the neural algorithm for style transfer from two perspectives: where to transfer and what to transfer. To specify where to transfer, we propose a simple yet effective strategy, named masking out, to constrain the transfer layout. To illustrate what to transfer, we define a new style feature by high-order statistics to better characterize content coherency. Without resorting to additional local matching or MRF models, the proposed method embeds the desired content information, either semantic-aware or saliency-aware, into the original framework seamlessly. Experimental results show that our method is applicable to various types of style transfers and can be extended to image inpainting.