Segmentation of Tippi + Consonant and Hoda + Consonant combination strokes in online Handwritten Gurmukhi Script Recognition

Abstract

There has been quiet a work done in the field of online handwriting recognition. It has been done for various scripts all over the world. In the online handwriting recognition for Indian scripts commendable work has been done as well. To further add to this extension of Indian scripts, here the work on Online Gurmukhi script handwritten word recognition is presented. When Gurmukhi script is written cursively, writer may write more than one character, vowels in a single stroke. Hence there is high chance of encountering combination of strokes written in a single stroke. These types of strokes are unrecognizable to the classifier. Segmentation algorithms are proposed that use the slope calculation method at every point and find candidate points for segmenting the stroke into individual basic strokes. The algorithm proposed here segments the Tippi+ consonant and Hoda+ consonant combination strokes. The algorithms demonstrated an accuracy of 96% in segmenting these stroke combinations when written in a single stroke. The first chapter gives the introduction to online Handwritten character Recognition. It put light on properties of the Gurmukhi script and the reasons that lead to difficulty in the recognition of online handwritten Gurmukhi script. The basic model followed for the process of handwritten character recognition id discussed. It discusses why there is a need for segmentation of strokes written by the writer in the case of online handwritten Gurmukhi characters. The second chapter discusses the earlier work that has already been done in this field. The previous works done by researchers in different scripts all over the world are discussed. The earlier work done on recognition in Gurmukhi script is also put light on. The third chapter explains the pre-processing steps applied on the strokes collected from the writer. Here the procedures of data compilation techniques, preprocessing steps (removing overlapping points, Normalization, centering, Missing point interpolation) are discussed in deep and the algorithms followed are also presented. The next chapter describes the problem that decreases the accuracy of recognition and the solution proposed for that problem. The model is proposed depicting when the segmentation process should be done. Its various phases are discussed in detail. And the final algorithm is presented. Finally the results of the implementation of the solution proposed are shown. The accuracy with which our system is able to successfully segment the strokes and the error rate is give. Conclusions and limitations of out proposed solution are also discussed.