A Novel Shape-Based Character Segmentation Method for Devanagari Script

This paper presents a new algorithm to extract shape-oriented feature vectors using pixel intensities from offline printed Devanagari script documents. Almost, all the characters of the script contain Shirorekha (header line) on the upper portion, which makes segmentation a difficult and complex problem. The problem gets more challenging when images are in multiple gray levels, skewed and noisy. A new fast and effective algorithm is designed using gradient structural information, and its performance is evaluated on a challenging dataset containing 80 printed documents consisting of around 87,000 characters. Experimental results show that the proposed algorithm has 98.56% accuracy, which is 02.66% higher than that reported in literature. Also, the proposed algorithm is time efficient and less complex in comparison with the existing methods.