New approach for dynamic textures discrimination

The aim of this paper is segmenting a sequence of images containing dynamic textures. The proposed method is based on means of features extracted from spatio-temporal cooccurrence matrices that characterize the textures themselves as well as their movements. Features with the highest discriminating power are selected according to a supervised scheme that permits to represent the dynamic textures in a relatively small dimensional space where they can be easily discriminated. Then the selected features are used in the segmentation phase. This approach has been applied in order to automatically identify the red alga (Gelidium Sesquipedale) on sequences of submarine images.