Invariant pattern recognition using SVDD-based associative memories

Abstract

Pattern recognition performances of a special gradient-type dynamical system are investigated. The system exhibits stable equilibrium points whose positions are defined by the minima of a data-dependent Lyapunov function constructed using the Support Vector Data Description (SVDD) algorithm. Invariance to standard geometric transformations is inferred by combining SVDD with the tangent distance (TD), which has superior recognition performances when compared to the Euclidean distance. Experimental results using the USPS handwritten characters database and the Olivetti face images database confirm the superiority of the proposed approach over existing solutions.