Statistical image object recognition using mixture densities

In this paper, we present a mixture density based approach to invariant ima ge object recognition. To allow for a reliable estimation of the mixture pa rameters, the dimensionality of the feature space is optionally reduced by applying a robust variant of linear discriminant analysis. Invariance to af fine transformations is achieved by incorporating invariant distance measur es such as tangent distance. We propose an approach to estimating covarianc e matrices with respect to image variabilities as well as a new approach to combined classification, called the virtual test sample method. Applicatio n of the proposed classifier to the well known US Postal Service handwritte n digits recognition task (USPS) yields an excellent error rate of 2.2%. We also propose a simple, but effective approach to compensate for local imag e transformations, which significantly increases the performance of tangent distance on a database of 1,617 medical radiographs taken from clinical da ily routine.