HMM based Two-Stage Classification Scheme to Improve Online Handwritten Kannada Numeral Recognition
M. Mahadeva Prasad, M. Sukumar · 2012
The paper proposes a HMM based two-stage classification scheme to recognize online handwritten Kannada numerals. Three sets of writer independent experiments are carried out with 1020 samples for training and 1016 samples for testing. First, experiments are carried out with a single-stage HMM classifier and from the confusion matrix, two majorly confused numeral pairs, and are identified. In the second set of experiments, by employing confusion pair-specific discriminating features, the individual classes of these confusion pairs are disambiguated by two separate two-class HMM based classifiers in the secondstage. In the third set of experiments, the training samples of each confusion pair are separately combined in the first-stage classifier to result in an eight-class problem. Accordingly, the first-stage classifier is trained to generate eight HMM models. Whenever the output of the primary classifier is one of these two combined classes, the corresponding test sample is passed on to the relevant two-class second-stage classifier to resolve the ambiguity. The Fourier transform of a set of features are used in the first-stage, whereas deviation features are employed in the secondstage classifiers. The respective average recognition accuracies of 96.5%, 98.5% and 98.2% are obtained with the first, second and third configurations of classifiers. The recognition performance of the two confusion pairs, and are increased by 5.8% and 3.6% respectively.