Offline handwritten Devanagari modified character recognition using convolutional neural network

In this work, two convolutional neural network (CNN)-based models are proposed for offline handwritten modified character recognition (e.g. [inline-graphic not available: see fulltext] ) in Devanagari script formed when a Devanagari consonant (e.g. [inline-graphic not available: see fulltext] ) is followed by a Devanagari vowel (e.g. [inline-graphic not available: see fulltext] ). The first model uses a single CNN architecture and the second model uses double-CNN architecture for the recognition of offline handwritten modified character. The double-CNN architecture performs better than single CNN architecture and uses a lesser number of output classes as compared with the actual existing classes of modified characters in Devanagari script. The recognition performance of these models is tested on Hindi consonants and Matras dataset with acceptable accuracy. The proposed CNN architecture yields better competitive results as compared with the traditional feature extraction (histogram of oriented gradients) and classification (support vector machine) techniques.