Image style transfer learning for style-strength control

Abstract

Image style transferring is a process of generating an output image in a target style from a given pair of content and target style images. Recently, a simple linear interpolation technique in encoded feature space has been employed in this process to generate output images of intermediate style because controlling the strength of style transferring effect is a key function of an image editing filter. However, this simple technique is effective in generating images of around full style but around zero style hence it cannot smoothly control the style transferring effect from an original content image to a fully stylised image. In this Letter, the authors tackle the missing work on stylestrength control from content reconstruction to full style generation. To deal with this problem, they propose to use additional unbiased training data and losses for a style transfer network to learn an unbiased regression between output style strength and style control parameter. Experimental results verified that the proposed method achieved a full range of style-strength control from zero style to full style with no additive complexity in image generating process.