The following technical report is available from http://aib.informatik.rwth-aachen.de/:
Jörg Dahmen Invariant Image Object Recognition using Gaussian Mixture Densities 2002-07
Abstract: In this work, a statistical image object recognition system is presented, which is based on the use of Gaussian mixture densities in the context of the Bayesian decision rule. Optionally, to reduce the number of free model parameters, a linear discriminant analysis is applied. This baseline system is then extended with respect to the incorporation of invariances. To do so, we start by suitably multiplying the available reference images. This idea is then applied to the observations to be classified, too, yielding the novel `Virtual Test Data' method, which has some desirable advantages over classical classifier combination approaches. Furthermore, global invariances are incorporated by using the so-called tangent distance. In this work, tangent distance is embedded into a statistical framework, which for instance leads to a modified, more reliable estimation of the mixture density parameters. Furthermore, tangent distance is extended to compensate not only for global, but also for local image transformations (`distorted tangent distanceŽ).
A large part of the experiments was performed on the well known US Postal Service standard corpus for handwritten digit recognition. Furthermore, the proposed classifier was successfully applied to the recognition of medical radiographs, red blood cells as well as to the Columbia University Object Image Library (COIL-20) and the Max-Planck Institute's Chair Image Database. The obtained error rate of 2.2% on the US Postal Service corpus is the best error rate published so far on this particular data set.
Regards, Volker