A multidimensional Modelling approach to texture classification and segmentation

The interest of an inverse filtering approach to picture modelling and recognition is emphasized. Following these lines, a multidimensional vector AR model is fitted to a reference region using a generalized Levinson procedure. The models of other regions are then used as inverse filters on the reference for classification. The approach is applied to natural pictures for recognition and segmentation using the textural features only. Some advantages of vector models over current bidimensional recursive models are indicated.