Cluster-based texture analysis

This paper presents a novel hierarchical cluster based texture model. The method first estimates a code-book for micro-texture in pixel neighbourhoods using a variant of Bezdek´s fuzzy c-mean clustering routine. The elements of the code-book are then used to “represent” the image. Larger scale textural variations are found by performing grouping over code-book elements in the “represented” image. At this stage a dispersed non-compact neighbourhood together with a c-medioid clustering method are used. The model falls into the class of structural texture models and depends on the multiple re-occurrence of micro-texture to function. The model is able to distinguish between Brodatz textures and detect “faults” in regular textures