Texture synthesis based on feature description

Patch-based texture synthesis techniques from a small sample texture generate a new texture image efficiently, but one significant problem is that these existing texture synthesis algorithms suffer from structure mismatching across patch boundaries due to inaccurate similarity measure. In this paper, we propose a desired distance metric to measure structural similarity when especially handling textures with salient features, which has the ability to find best matches during the patch-based sampling. To do this, we introduce the concept of "layers" to describe the correlation of pixels through computing each pixel weight, and construct the orientation fields to capture the inherent direction and variation of texture features through computing the dominant local orientation angle and the degree of anisotropy of each pixel. Our experimental results demonstrate that the proposed method can significantly alleviate structure discontinuities in synthesized textures.