Statistical Image Object Recognition
Daniel Keysers

Introduction
In this project, statistical pattern recognition approaches are used to classify objects which are present in a given image. In the Lehrstuhl für Informatik VI classifier, classification is performed using the Bayesian decision rule. Here, the distribution of the reference images is modelled using Gaussian mixture or kernel densities, where the parameters of the distribution are usually estimated using the Expectation-Maximization algorithm.

After the successful application of discriminative training procedures, we studied methods to make the recognition process invariant against certain types of image transformations such as rotation, size scaling and translation. To this purpose, the so called tangent approximation was used in the context of Gaussian mixtures. By this method the error rate on the US Postal Service handwritten digits recognition task could be reduced from 4.2% to 2.4% .

Examples from the USPS Database

Overview of Research Activities
In this group, algorithms coming from statistical pattern recognition are used for image processing and recognition. Here the goal is to classify image data reliably and efficiently.

To achieve this goal, the observations are treated as resulting from an unknown probability density function. This distribution must be described choosing adequate models and estimating the parameters in a training phase. In the implemented classifiers for example Gaussian mixture densities are chosen and the parameters are estimated using the expectation-maximization algorithm. The trained models are then used in the recognition process, in which the unknown class is determined using the Bayesian decision rule, choosing the class with the highest probability. Because of the high dimensionality of the feature vectors that usually arise in image processing these may be reduced using for example a linear discriminant analysis.

The knowledge obtained from the classification of single objects will be used to also classify more complex scenes. Here, the Bayesian decision rule is extended in order to allow the classification of multiple objects in an image also taking into account varying background. The resulting search space must be limited using appropriate methods. The developed methods are planned to be used for example for object-based image retrieval.

At i6 typical application areas are handwritten optical character recognition or the classification of red blood cells. Other databases used include medical images from the IRMA-project of the RWTH Aachen University of Technology and the Columbia University Object Image Libarary.

Special emphasis is put on the modelling of variability and incorporation of invariances into the statistical model. This can be achieved using for example so-called tangent vectors that can be integrated into the statistical framework efficiently. For local invariances the usage of the image distortion model is also helpful.

Within this group we continuously offer projects for students for Diploma theses or research projects (Diplom-/ Studienarbeiten). Interested students please contact

Dipl.-Inform. Daniel Keysers
Lehrstuhl für Informatik VI
Ahornstrasse 55
52056 Aachen
Tel.: +49-241-80-21610
keysers@informatik.rwth-aachen.de

Some Related Publications:

D. Keysers, W. Macherey, J. Dahmen, and H. Ney. Learning of Variability for Invariant Statistical Pattern Recognition. In ECML 2001, 12th European Conference on Machine Learning. Freiburg, Germany, pages 263-275, Volume 2167 of Lecture Notes in Computer Science, Springer, September 2001. (abstract, .ps.gz, , .pdf)

J. Dahmen, D. Keysers, and H. Ney. Combined Classification of Handwritten Digits using the 'Virtual Test Sample Method'. MCS 2001, 2nd International Workshop on Multiple Classifier Systems. Cambridge, UK, pages 109-118, Volume LNCS 2096 of Lecture Notes in Computer Science, Springer, July 2001. (abstract, )

J. Dahmen, D. Keysers, H. Ney, and M. O. Güld. Statistical Image Object Recognition using Mixture Densities. Mathematics and Image Analysis Conference, Paris, France, September 2000. Published in Journal of Mathematical Imaging and Vision, Volume 14, Number 3, pages 285-296, May 2001. Kluwer

D. Keysers, J. Dahmen, T. Theiner, H. Ney, "Experiments with an Extended Tangent Distance". 15th International Conference on Pattern Recognition, pp. 38-42, Barcelona, Spain, September 2000.

J. Dahmen, D. Keysers, M. Güld, H. Ney, "Invariant Image Object Recognition using Mixture Densities", 15th International Conference on Pattern Recognition, pp. 614-617, Barcelona, Spain, September 2000.

D. Keysers, J. Dahmen, H. Ney, "A Probabilistic View on Tangent Distance", 22. DAGM Symposium Mustererkennung, pp. 107-114, Kiel, Germany, September 2000.

J. Dahmen, D. Keysers, Michael Pitz, H. Ney, "Structured Covariance Matrices for Image Object Recognition", 22. DAGM Symposium Mustererkennung, pp. 99-106, Kiel, Germany, September 2000.

J. Dahmen, T. Theiner, D. Keysers, H. Ney, T. Lehmann, and B. Wein. "Classification of Radiographs in the `Image Retrieval in Medical Applications' System (IRMA)", 6th International RIAO Conference on Content-Based Multimedia Information Access, Paris, France, pages 551-566, April 2000.

J. Dahmen, K. Beulen, M. Güld, H. Ney, "A Mixture Density Based Approach to Object Recognition for Image Retrieval",
6th International RIAO Conference on Content-Based Multimedia Information Access, pp. 1632-1647, Paris, France, April 2000.

J. Dahmen, J. Hektor, R. Perrey, "Automatic Classification of Red Blood Cells using Gaussian Mixture Densities", Bildverarbeitung für die Medizin 2000, pp. 331-335, Munich, Germany, March 2000.

J. Dahmen, R. Schlüter, H. Ney. "Discriminative Training of Gaussian Mixtures for Image Object Recognition", 21. DAGM Symposium Mustererkennung, W. Förstner, J. Buhmann, A. Faber, P. Faber (eds.), pp. 205-212, Bonn, Germany, September 1999.