No abstract

No abstract

Abstract. Detection of regions similar to a single given texture in arbitrary colour images is difficult for conventional supervised or unsupervised segmentation techniques. We introduce a novel partially supervised algorithm that solves this problem using similarity between local statistics on different levels of pyramidal representations of the texture and the image. Most characteristic statistics for the texture are estimated in accord with a generic Gibbs random field model with spatially homogeneous pairwise pixel interactions. Empirical distributions of the self-similarity values for the texture itself are used to separate the desired texture from an arbitrary background. Experiments with different images, including aerial images of the Earth's surface, show this algorithm effectively detects regions with spatially homogeneous and weakly homogeneous textures.