On the Performance of Devnagari Handwritten Character Recognition
Rajiv Kumar, Kiran Kumar Ravulakollu · 2014
This paper presents the offline handwritten character recognition for Devnagari, a major script of India. The main objective of this work is to develop a handwritten dataset (CPAR-2012) for Devnagari character and further develop a character recognition scheme for benchmark study. The present dataset is a new development in Devnagari optical document recognition. The dataset includes 78,400 samples collected from 2,000 heterogeneous strata of Hindi speaking persons. These dataset is further divided into 49,000 as training set and 29,400 as test set. The evaluated feature extraction includes: direct pixel, image zoning, wavelet transformation and Gaussian image transformation techniques. These features were classified by using KNN and neural network classifier. The experiment shows that Gaussian image transformation (level 1) using KNN classifier has achieved highest recognition 72.18 % than other feature extraction methods. Further classification result obtained from KNN classifier were combined, the combined result shows 84.03 % recognition accuracy with expense of 5.3 % rejection. Based on this result some shape similar character zones in Devnagari characters are highlighted in this paper.