Performance Comparison of Several Feature Selection Techniques for Offline Handwritten Character Recognition

This paper presents a performance comparison of various feature selection techniques for offline handwritten Gurmukhi character recognition. Research on offline handwritten character recognition of Gurmukhi script is very difficult due to the complex structural properties of the script that are not matter-of-fact in most other scripts. Gurmukhi is the script used for writing the Punjabi language, which is the official language of Punjab state in India. We have presented a feature extraction technique for offline handwritten Gurmukhi character recognition based on the boundary extent of the character image and used various feature selection techniques, to reduce the dimensionality of feature vectors. We have also compared their recognition performances using two different classifiers, namely, Nearest Neighbours (NN) and Support Vector Machine (SVM) with linear kernel. Different classification schemes measures are used for the performance analysis of different feature selection techniques. Results obtained using presented feature extraction technique show that Chi Squared Attribute (CSA) feature selection technique performs better than other feature selection techniques using NN and SVM with linear kernel classifier for character recognition. In this work, we have obtained zone wise maximum recognition accuracy of 88.3%, 95.2% and 91.3% for upper zone, middle zone and lower zone of Gurmukhi script, respectively.