A Method of Neural Style Transfer for Images with Artistic Characters

Abstract

There is a technique called neural style transfer for recognizing complicated and unobvious relationships between the contents of an image and its painting style by utilizing a deep learning technique. It can automatically generate artistic images based on the acquired knowledge. The method is effective for processing some kinds of image contents such as landscape images. When a target image includes characters such as logos, it tries to transfer the whole image style including the characters. The result becomes a problematic one because the characters tend to be washed out in the transferred image. In this paper, we propose a method for producing better results in the style transfer on images with characters. We devise a method for clearly preserving the forms of the characters included in an input image when fusing the learned style with the characters artistically. We confirm the effectiveness of the proposed method by comparing images created by our proposed method and a conventional method. Then we discuss some issues for further improvements in our future study.