Multi-scale texture classification from generalized locally orderless images

Locally orderless images are families of three intertwined scale spaces that describe local histograms. We generalize locally orderless images by considering local histograms of a collection of filtered versions of the image, and by extending them to joint probability distributions. These constructions can be used to derive texture features and are shown to be a more general description of two established texture classification methods, viz., filter bank methods and cooccurrence matrices. Because all scale parameters are stated explicitly in this formulation, multi-resolution feature sets can be extracted in a systematic way. This includes new types of multi-resolution analysis, not only based on the spatial scale, but on the window size and intensity scale as well. Each multi-resolution approach improves texture classification performance, the best result being obtained if a multi-resolution approach for all scale parameters is used. This is demonstrated in experiments on a large data set of 1152 images for 72 texture classes