Combining HMM and SVM based Stroke Classifiers for Online Assamese Handwritten Character Recognition

Hidden Markov Models (HMMs) and Support Vector Machine (SVM) based classifiers are commonly used in the field of handwriting recognition. In this paper we investigate a technique of recognizing Assamese handwritten characters using HMMs and SVM stroke classifiers in conjunction to each other. The two classifiers are separately trained on same stroke dataset with same set of features. The top-N class labels from the HMM classifier are selected and the search space of the SVM classifier is reduced by inspecting the SVM scores for the selected N classes only. The class with highest SVM score among these N classes is the predicted class. However the confidence score from the classifier is low for a predicted class if there exist some confusion with similar classes. In such cases a fall-back option to consider the decision of the HMM classifier is introduced depending on the confidence score from the SVM classifier. In this way we select decision from one of the classifier to increase the stroke recognition rate. We have observed that better recognition performance can be achieved by proposed method. Finally for recognition of a test character, the recognized stroke set from the combined classifier is matched with the stroke sequences in a lookup table or reference set. The experiments are performed in a large number of handwritten Assamese characters collected from 100 native writers.