Narrow Results

This paper presents a hybrid approach to recognition of handwritten basic characters of Bangla, the official script of Bangladesh and second most popular script of India. This is a 50 class recognition problem and the proposed recognition approach consists of two stages. In the first stage, a shape feature vector computed from two-directional-view-based strokes of an input character image is used by a hidden Markov model (HMM). This HMM defines its states in a data-driven or adaptive approach. The statistical distribution of the shapes of strokes present in the available training database is modelled as a mixture distribution and each component is a state of the present HMM. The confusion matrix of the training set provided by the HMM classifier of the first stage is analyzed and eight smaller groups of Bangla basic characters are identified within which misclassifications are significant. The second stage of the present recognition approach implements a combination of three multilayer perceptron (MLP) classifiers within each of the above groups of characters. Representations of a character image at multiple resolutions based on a wavelet transform are used as inputs to these MLPs. This two stage recognition scheme has been trained and tested on a recently developed large database of representative samples of handwritten Bangla basic characters and obtained 93.19% and 90.42% average recognition accuracies on its training and test sets respectively.