Multi-Scale Label-Map Extraction for Texture Synthesis

摘要

Texture synthesis is a well-established area, with many importantrnapplications in computer graphics and vision. However, despiterntheir success, synthesis techniques are not used widely in practicernbecause the creation of good exemplars remains challenging andrnextremely tedious. In this paper, we introduce an unsupervisedrnmethod for analyzing texture content across multiple scales thatrnautomatically extracts good exemplars from natural images. Unlikernexisting methods, which require extensive manual tuning, our method is fully automatic. This allows the user to focus on usingrntexture palettes derived from their own images, rather than on manualrninteractions dictated by the needs of an underlying algorithm.rnMost natural textures exhibit patterns at multiple scales that mayrnvary according to the location (non-stationarity). To handle suchrntextures many synthesis algorithms rely on an analysis of the inputrnand a guidance of the synthesis. Our new analysis is based on arnlabeling of texture patterns that is both (ⅰ) multi-scale and (ⅱ) unsupervisedrn– that is, patterns are labeled at multiple scales, and thernscales and the number of labeled clusters are selected automatically.rnOur method works in two stages. The first builds a hierarchicalrnextension of superpixels and the second labels the superpixelsrnbased on random walk in a graph of similarity between superpixelsrnand a nonnegative matrix factorization. Our label-maps providerndescriptors for pixels and regions that benefit state-of-the-art texturernsynthesis algorithms. We show several applications includingrnguidance of non-stationary synthesis, content selection and texturernpainting. Our method is designed to treat large inputs and can scalernto many megapixels. In addition to traditional exemplar inputs, ourrnmethod can also handle natural images containing different texturedrnregions.