Exploring Patch Similarity in an Image
Jose-Luis Lisani, Jean-Michel Morel
⚠ This is a preprint. It may change before it is accepted for publication.


This article describes an experimental procedure to analyze the self-similarity concept in natural images and to explore the Gaussianity of groups of similar patches extracted from a single image. The self-similarity assumption takes it that most image patches of a sufficient size are repeated, of course not identically, but with small variations. The procedure proposed in this paper and implemented in the accompanying online demo permits to explore and visualize these clusters of similar patches in a given image. Thanks to it a user can select a patch in an image, group all patches similar to it up to a translation, or to an isometry, apply PCA to the group, make visual tests about the Gaussianity of the set of patches, and finally apply EM to the set to see if it is a mixture of Gaussians.