Style Transfer Based on Style Primitive Discovery

Hao Wu, Zhengxing Sun, Shuang Wang, Weihang Yuan, Hui-Hsia Chen

Abstract: Example-based stylization provides a direct way of making artistic effects for images. However existing methods are not suitable for artistic applications like Chinese embroidery. In this paper, we propose an example-based non-rigid image stylization method tailored for Chinese embroidery art. To this aim, a novel style transfer framework is presented, which works by using different aggregation patterns, i.e. regular primitive and stochastic primitive. We find that these two patterns are surprisingly effective in embroidery description. Specifically, we first extract these two style primitives from an example image according to the directionality and orientation. Then we employ a primitive selection algorithm to filter defected primitives. After that, we employ a sparse representation-based style transfer method, to synthesize the final result. In the experiments, the synthesis results show that our framework is superior to state-of-the-art methods and performs more efficient on large resolution images.