Devanagari Online Handwritten Character Recognition

Abstract

In this thesis, a classifier based on local sub-unit level and global character level representations of a character, using stroke direction and order variations independent features, is developed for recognition of Devanagari online handwritten characters. It is shown that online character corresponding to Devanagari ideal character can be analyzed and uniquely represented in terms of homogeneous sub-structures called the sub-units. These sub-units can be extracted using direction property of online strokes in an ideal character. A method for extraction of sub-units from a handwritten character is developed, such that the extracted sub-units are similar to the sub-units of the corresponding ideal character. Features are developed that are independent of variations in order and direction of strokes in characters. The features are called histograms of points, orientations, and dynamics of orientations (HPOD) features. The method for extraction of these features spatially maps co-ordinates of points and orientations and dynamics of orientations of strokes at these points. Histograms of these mapped features are computed in di erent regions into which the spatial map is divided. HPOD features extracted from the sub-units represent the character locally; and those extracted from the character as a whole represent it globally. A classifier is developed that models handwritten characters in terms of the joint distribution of the local and global HPOD features of the characters and the number of sub-units in the characters. The classifier uses latent variables to model the structure of the the sub-units. The parameters of the model are estimated using the maximum likelihood method. The use of HPOD features and the assumption of independent generation of the sub-units given the number of sub-units, make the classifier independent of variations in the direction and order of strokes in characters. This sub-unit based classifier is called SUB classifier. Datasets for training and testing the classifiers consist of handwritten samples of Devanagari vowels, consonants, half consonants, nasalization sign, vowel omission sign, vowel signs, consonant with vowel sign, conjuncts, consonant clusters, and three more short strokes with di erent shapes. In all, there are 96 di erent characters or symbols that have been considered for recognition. The average number of samples per character class in the training and the test sets are, respectively, 133 and 29. The smallest and the largest dimensions of the extracted feature vectors are, respectively, 258 and 786. Since the size of the training set per class is not large compared to the dimension of the extracted feature vectors, the training set is small from the perspective of training any classifier. classifiers that can be trained on a small data set are considered for performance comparison with the developed classifier. Second order statistics (SOS), sub-space (SS), Fisher discriminant (FD), feedforward neural network (FNN), and support vector machines (SVM) are the other classifiers considered that are trained with the other features like spatio-temporal (ST), discrete Fourier transform (DFT), discrete cosine transform (DCT), discrete wavelet transform (DWT), spatial (SP), and histograms of oriented gradients (HOG) features extracted from the samples of the training set. These classifiers are tested with these features extracted from the samples of the test set. SVM classifier trained with DFT features has the highest accuracy of 90.2% among the accuracies of the other classifiers trained with the other features extracted from the test set. The accuracy of SUB classifier trained with HPOD features is 92.9% on the test set which is the highest among the accuracies of all the classifiers. The accuracies of the classifiers SOS, SS, FD, FNN, and SVM increase when trained with HPOD features. The accuracy of SVM classifier trained with HPOD features is 92.9%, which is the highest among the accuracies of the other classifiers trained with HPOD features. SUB classifier using HPOD features has the highest accuracy among the considered classifiers trained with the considered features on the same training set and tested on the same test set. The better character discriminative capability of the designed HPOD features is re ected by the increase in the accuracies of the other classifiers when trained with these features