DFT Based Feature Extraction Technique for Recognition of Online Handwritten Gurmukhi Strokes using SVM

Abstract

Small devices face some physical problems for having a keyboard. So, there is a need for providing human and machine communication. Communication can be done either through speech or through writing. Natural handwriting is one way of exchanging information. Online handwriting recognition system can be used as a medium for providing a natural way of communication between user and computer. Since pen based devices are emerging rapidly, presence of online handwriting recognition feature in such devices is quite useful. Research work presented in this thesis focused on recognition of online handwritten Gurmukhi strokes based on Discrete Fourier Transform features using Support Vector Machine. The proposed method works in two stages. In first stage, a DFT based feature extraction method is applied on the preprocessed strokes and in second stage, classification of strokes is done using SVM classifier. We have considered 86 stroke classes of Gurmukhi script in this work and for each class 75-100 variations are considered. Lastly, after testing of the proposed method on a data set of 8408 stroke samples, a recognition accuracy of 91.7 % has been achieved when 11-fold cross-validation approach in LibSVM with RBF kernel is used.