Character recognition based on trellis diagrams

Many papers on pattern recognition have been published in the last decades but character recognition is still part of current research. However, one main topic are classifiers that can be easily augmented by new training data or even new classes. Furthermore, the classifiers have to have a certain robustness with respect to noise, i.e., the recognition rate must not be significantly affected by the presence of noise in the character images. For this reason a new classifier approach is introduced, which is based on trellis diagrams and thus similar to a Viterbi decoder known from communication systems. The training as well as the classification procedure are discussed in detail. Additionally, to show the competitiveness the performance is compared with already existing classifiers on a character dataset with and without noise.