Abstract

The main goal of pattern recognition is to achieve a high recognition performance. One popular way to manage this is to develop and optimize a feature for the particular task and to train and optimize a matching classifier. But sometimes it is not so clear which feature or classifier approach are the most qualified ones. Therefore, another option is to create several classifiers using different feature views on the data and then combine the outputs of the individual classifiers in an appropriate way. A classifier constructed in such way is called multiple classifier system (MCS). In order to obtain a good MCS performance the individual classifiers, that are to be fused, do not only have to show reasonable individual accuracies, but should also be diverse. In particular if a sample is classified wrongly by a certain classifier, the other classifiers should not agree on a false label to perceive the ability to correct it by combining the classifiers.

In this talk MCS will be explained and demonstrated by means of two different applications: firstly an approach to discover events in data recorded with an electroencephalograph (EEG) is described. For this study the “Machine Learning for Signal Processing 2010 Competition: Mind Reading” dataset is used. Visual stimuli were presented to a test person by following a typical oddball paradigm: the non target (background) type stimuli were presented very frequently, whereas the targets were displayed very rarely which leads to prominent event related potentials in the EEG. The challenge of this dataset is the classification of stimuli by solely analyzing the EEG recordings. Secondly, a multiple classifier approach for the recognition of facial expressions in image sequences is presented. For this second evaluation, the well known “Cohn-Kanade Comprehensive Database for Facial Expression Analysis” is used.