Spatially Aware Style Transfer

The task of image style transfer is to automatically redraw (using neural networks) an image with some content (for example, a family photo) in the style set by another image (for example, a Van Gogh painting), which finds applications in advertising, design, entertainment, and other fields. Common stylization algorithms extract and apply the style evenly, which limits the expressiveness of the stylized result and does not correspond to the real work of artists who use the style differently for different objects in the image, such as a portrait of a person and background objects. The paper proposes an improved style transfer algorithm due to non-uniform styling: style and content images are divided into regions, then each content region is styled with a style from the most appropriate areas of the style image. Improvements of the proposed method compared to existing analogues are shown at a qualitative level, as well as by polling respondents who were required to choose the best stylization among stylizations using different methods in random order.