EFFECT OF NON-IMAGE FEATURES ON RECOGNITION OF HANDWRITTEN ALPHA-NUMERIC CHARACTERS

Abstract

Handwritten character recognition has applications in several industries such as Banking for reading of cheques and Libraries/ National archives for digital searchable storage of historic texts. The main feature typically used for the recognition task is the character image. However, there are other possible features such as the hand (left or right) used by author, number of strokes and other geometric features that can be captured when writing on digital devices.  This paper investigates the effect of using some non-image features on the recognition rate of three classifiers: Instance Based Learner (IBk), Support Vector Machines (SVM) and the Multilayer Perceptron (MLP) Neural Network for singly-written alpha-numeric character recognition. Our experiments were conducted using the WEKA machine learning tool on offline and online handwritten acquired locally. A percentage split (66%-34% train-test) evaluation methodology was adopted with the classification accuracy measured. Results indicate that non-image additional features improved the accuracy across the three classifiers for the online and offline character datasets. However, this improvement was not statistically significant. SVM gave the best accuracy for the online dataset while IBk performed better than the other two classifiers for the offline dataset. We intend to investigate the effect of non-image features at other levels of text granularity such as words and sentences