Content-Aware Texture Synthesis

Existing example-based texture synthesis techniques are inherently unadapted to textures consisting of a set of randomly disposed, individually discernible shapes. Local methods striving at pixel-based discontinuity reduction hardly preserve input's long-range structures. Alternatively, research built upon the supposed respect by the input's features of given placement rules are too restrictive to be straightly extended to stochastic arrangements. In this paper we present a new method for analyzing and resynthesizing such arrangements. Our objective is to acquire their constitutive shapes to enable structure-aware resynthesis. What characterizes such shapes is their repetition throughout the input. We exploit this trait by recording recurrences of visually similar neighborhoods which are later extended to regions. We bring those together to compute the input's coverage map and extract final repetitive shapes. By directly manipulating shapes, resynthesis can be enriched with high-level information unavailable in pixel-based approaches. We gather statistics on their placement and appearance variations and use those to produce new images. To achieve this, we draw inspiration and improve techniques for capturing element arrangements, techniques once limited to vectorized NPR primitives.