A dominant points-based feature extraction approach to recognize online handwritten strokes

The computation of features is an integral part of handwriting recognition as identification of correct features is essential for efficient data representation and benchmarked recognition. In context of handwriting recognition, the aim of feature extraction is to find out the certain properties of a handwritten stroke that best describe the class of a stroke and makes it distinguishable from other stroke classes. The present work proposes a novel approach for feature extraction based on dominant points in online handwritten strokes. The proposed scheme finds the curve directions between the consecutive dominant points of the stroke and prepares the fixed length feature vector for a handwritten stroke, as the input of fixed length feature vector is used for statistical recognition techniques such as support vector machines and hidden Markov models. Therefore, it avoids the limitation of original Ramer–Douglas–Peucker technique that gives variable number of dominant points. Our approach recognizes online handwritten strokes without huge preprocessing and with a smaller length of feature vector. The proposed technique can be used for different writing styles and different character sets. The efficiency of proposed approach is evaluated on two different datasets as inhouse dataset of 39,200 strokes collected from 100 users and UNIPEN dataset, where consistent and reliable recognition accuracy has been attained for both datasets. The major objective of present work is to propose a dominant points-based script-independent feature extraction technique for online handwriting that is suitable for real-life applications.