Arbitrary Style Transfer with Semantic Content Enhancement

Abstract

Arbitrary style transfer is an import topic which changes the style of a source image according to a reference one. It is useful for artistic creation and intelligent imaging applications. The main challenge of the style transfer is that it is difficult to balance the semantic feature transformation and original semantic content. In this paper, we introduce a semantic content enhancement module to mitigate the affect of color distribution and semantic feature transformation for the style transfer while keeping the original semantic structure as much as possible. Meanwhile, we also introduce a channel attention module to enhance the style features by fusing with the style attention network. With the enhancement of both features, our network achieves excellent result that balances original semantic structure and transfer stylized visualization. In addition, we also migrate the algorithm to 3D space and it also performs stably for 3D scene-based style transfer. Experiments show that our method can handle various style transfer tasks.