Saliency-Guided Image Style Transfer

Abstract

In this paper, we propose an automatic saliency-guide style transfer method, which deploys style transfer to the salient object in an image. Our method can generate a new artwork form of image, which aggregates the characteristics of real-world images and artistic images. First, the existing style transfer method is utilized to combine the content of an arbitrary real-world image with the appearance of a well-known artwork, and to produce the stylized image with high perceptual quality. Then, the saliency map of real-world image is used to fuse the real-world image and the stylized image. In this way, the fusion image not only contains the style of artwork but also highlights the salient object. Finally, the saliency-based refinement is employed to improve the quality of object boundaries, and to enable the stylized object blends into the surroundings naturally. Experimental results demonstrate the effectiveness of the proposed saliency-guided image style transfer method.