Writer Identification System for Handwritten Gurmukhi Characters: Study of Different Feature-Classifier Combinations

In this paper, we are exploring various features and classifiers for writer identification in light of Gurmukhi text handwriting. The identification of the writers based on a piece of handwriting is a challenging task for pattern recognition. The writer identification framework proposed in this paper includes diverse stages like image preprocessing, feature extraction, training, and classification. The framework first prepares a skeleton of the character so that meaningful data about the handwriting of writers can be extracted. The feature extraction stage incorporates various plans, namely, zoning, diagonal, transition, intersection and open end points, centroid, the horizontal peak extent, the vertical peak extent, parabola curve fitting, and power curve fitting based features. In order to assess the prominence of these features, we have used four classification techniques, namely, Naive Bayes, Decision Tree, Random Forest and AdaBoostM1. For experimental results, we have collected 49,000 samples from 70 different writers. In this work, maximum accuracy of 81.75% has been obtained with centroid features and AdaBoostM1 classifier.