Neural Style Transfer with Content Discrimination

Neural style transfer, which creates fantastically stylized images by separating and recombining the content and style, has received much attention in both academic and industrial applications. Thus, how to separate and recombine the content and style is of critical importance. However, most of existing neural style transfer algorithms attempt to stylize the content images uniformly without considering the variations of content details, thus sometimes producing visually unpleasing images. To address this problem, we propose a novel content-aware neural style transfer algorithm by taking account of variations of image contents (e.g., smooth, salient regions). To this end, our method first divides the content image into different regions based on salient information and then stylizes different regions discriminatively. Such a content-aware way of transfer retains more identifiability of the subject in the content image resulting in higher appreciable value in image quality than general transfer. Experiments demonstrate the effectiveness of our method.