Multiple classifier systems in offline cursive handwriting recognition

The thesis addresses the application of multiple classifier systems (MCS) methods in the domain of handwriting recognition. To evaluate the MCS methods two different state-of-the-art handwritten word recognizers are used. Both recognizers are based on Hidden Markov Models (HMMs) and process handwritten words in the same manner: First the word image is normalized in respect to slant, skew and height. Then a sequence of feature vectors is extracted from the image. Finally the HMMs are used to determine the most likely word class from a set of candidate word classes. As the performance of every MCS method depends on the performance of the classifiers it is applied to, the optimization of the recognizers is also addressed in the thesis. There are mainly two kind of MCS methods: The ensemble methods which create several classifiers out of one base classifier and the combination schemes which combine the output of several classifiers to a single output. Both kind of methods are addressed in the thesis. Also the optimal values of the parameters of the MCS methods, such as the number of classifiers created by the ensemble methods, are evaluated in the thesis. There are two objectives of the thesis. The first objective is to the design new MCS methods and the second objective is to evaluate some classic and the new MCS methods in large scale experiments. Four new categories of ensemble methods and two new combination schemes are proposed in the thesis. In addition the adaption of some classic combination schemes for the use with HMM classifiers is addressed. All methods are tested in a handwritten word recognition task where the recognizer, given an image of a word, has to choose a word from 3997 possible word classes.