ON STUDYING HANDWRITTEN DEVANAGARI WORD RECOGNITION SYSTEM BASED ON GRADIENT AND STRUCTURAL FEATURES
*Sukhjinder Singh 1and Naresh Kumar Garg2
*1Department of Electronics & Communication Engineering, GianiZail Singh Campus College of Engineering & Technology, Maharaja Ranjit Singh Punjab Technical University, Bathinda-151001, Punjab (India); email@example.com ; firstname.lastname@example.org
2Department of Computer Science & Engineering, GianiZail Singh Campus College of Engineering & Technology, Maharaja Ranjit Singh Punjab Technical University, Bathinda-151001, Punjab (India);;email@example.com
Abstract: This paper introduces a method for the recognition/identification of the offline handwritten Devanagari words, which has a variety of pattern recognition applications includingsuch as cheque reading, airline ticket readers, bill processing systems, handwritten address interpretation, signboard translation and postal automation. The proposed framework uses a holistic approach for the recognition/identification of Devanagari handwritten words. It considers, a complete word as an individual entity for recognition i.e. without using segmentation. Gradient and structural features based on contour-directional histogram are extracted from the handwritten word images. Three different classifiers namely SVM (Support Vector Machine), NB (Naive Bayes) and XGBoost(eXtreme Gradient Boosting) are used for the recognition tasks. Experiments are also carried out using combined feature vectors resulted from gradient and structural features as input to various classifiers. The framework is evaluated on the corpus of 20,000 words of 50-different town names handwritten in Devanagari script. The presented work is subjected to performance evaluation in terms of PR (Precision), FAR (False Acceptance Rate), FRR (False Rejection Rate), and RA (Recognition Accuracy). It has gathered from experimental work that combination of feature vectors resulted from gradient and structural features along XGBoost classifier perform better as compared with individual features itself.
Keywords: Handwritten word recognition; holistic approach; feature extraction; classification