Devanagari offline handwritten numeral and character recognition using multiple features and neural network classifier
Vikas J. Dongre, Vijay H. Mankar · International Conference on Computing for Sustainable Global Development · 2015
This paper presents an attempt to solve the challenging problem of Devanagari numeral and character recognition. It uses structural and geometric features to represent the Devanagari numerals and characters. Each image is zoned in 9 blocks and 8 structural features are extracted from each block. Similarly 9 global geometric features are extracted. These 81 features are used for representing the image. Multilayer perceptron neural network (MLP-NN) is used for classification. 3000 handwritten samples of Devanagari numerals and 5375 handwritten samples of Devanagari alphabetic characters are used for training and testing. Experimental results show 93.17% recognition accuracy using 40 hidden neurons for numerals and 82.7% recognition accuracy using 60 hidden neurons for characters. Fivefold cross validation is used for verifying the results.