Robust features for textures in additive noise C.Ottonello l, S.Pagnan 2 and V. Murino 1 1Dip. di Ingegneria Biofisica ed Elettronica-University of Genoa Via alrOpera Pia 11A, 16145 Genova, Italy 2Istituto di Automazione Navale- National Research Council of Italy Torre di Francia, Via De Marini 1, 16149 Genova, Italy Abstract. The paper describes a method for texture classification in noise by using third- order cumulants as discriminating features. The problem is formulated as a test on K hypotheses and solved by a Maximum Likelihood (ML) criterium applied in the third-order cumulant domain. Since in the case of image processing complete third-order cumulant computation is not feasible, we reduced the estimation to a limited number of cumulant slices and lags. This reduction makes the classification algorithm suboptimal. Thus, a criterion for the choice of cumulant samples to be computed is introduced in order to guarantee the selection of those lags which better identify the different textures in the training phase of the classifier. Experimental tests are carried out to evaluate third-order cumulant performances on noisy textures and the importance of lags selection. 1 Introduction The aim of texture classification algorithms is to produce a set of measures that make it possible to discriminate between different classes of textures so that each class may be described by parameters that can be used by a segmentation algorithm to partition an image into homogeneous regions. When images are affected by noise, this purpose can be strongly compromised and it is necessary to produce robust measures for each class of texture. In the paper, a Maximum Likelihood (ML) classifier working in the domain of Higher Order Statistics (HOS) [1] is proposed; in particular, third-order cumulants are estimated as texture features, thanks to their insensitivity to symmetric and independent, identically distributed (i.i.d.) noises. In order to reduce computation efforts and to determine an appropriate set of cumulant samples to be included in the feature vector, the number of outliers for each class is evaluated during the training phase. Cumulant samples inconsistent with the expectation produced by each class population are ranked out by assigning them small efficiency weights. In section 2, the classifier scheme is presented, and, in subsection 2.1, the adopted criterion for lag selection is discussed. Section 3 describes the experiments performed on natural textures corrupted by i.i.d, as well as coloured Gaussian noise; in this section, a comparison with an autocorrelation based classifier is also made. Some conclusions are carried out in section 4.