Mathware & Soft Computing 10 (2003) 155-168 Innovative Applications of Associative Morphological Memories for Image Processing and Pattern Recognition M. Gra~ na 1 , P. Sussner 2 , G. Ritter 3 1 Dept. CCIA, UPV/EHU, San Sebastian, Spain ccpgrrom@si.ehu.es 2 Institute of Math.,Stat. and Sci. Comp. State University of Campinas, Brazil 3 Center for Comp. Vision and Visual. University of Florida A bst ract Morphological Associat ive Memories have been proposed for some image denoi sing applicat ions. T hey can be appl ied t o ot her l ess r est ri ct ed domai ns, like image retrieval and hyperspectral image unsupervised segmentation. In this paper we present these applications. In both cases the key idea is that Autoassociative Morphological Memories selective sensitivity to erosive and dilative noise can be applied to detect the morphological independence be- t ween pat t er ns. Linear unmixing based on the sets of morphological inde- pendent pat t erns de¯ ne a feat ure extract ion process t hat is t he basis for t he image processing applications. We discuss some experimental results on the ¯ sh shape data base and on a synthetic hyperspectral image, including the comparison with other linear featureextraction algorithms (ICA and CCA). 1 Introduction Linear feature extraction algorithms, like Principal Component Analysis (PCA) [2], Linear Discriminant Analysis (LDA) [2], Independent Component Analysis (ICA) [7] are de¯ned as a linear transformation that minimizes some criterion function, like the mean square error (PCA), a class separability criterion (LDA) or an independence criterion (ICA). The approach we take is to try to characterize the data by a convex region that encloses them or most of them. The features extracted are the relative coordinates of the data points in this region. In other words the result of the linear unmixing relative to the vertices of this convex region. Therefore the dimensionality reduction depends on the degree of detail of the de¯nition of this convex region: the number of vertices that describe it. Depending on the application, the meaning of these vertices varies. In hyperspectral image processing 155