A low-cost hybrid handwritten Devanagari character classifier

Optical character recognition (OCR) of handwritten documents is a popular field of research due to its application in real life. It is a way of digitizing the documents. Though digitization of the handwritten documents is a complex research problem due to the challenges, such as skewed writing and various writing styles associated with it, still it has a widespread utility in social life as well as preserving biblical information. Starting from banking, hospitals, education and government offices OCR is a solution to save human time and effort. Classification is the most important and complex step, which has the maximum contribution on the accuracy of the OCR system. In this article, we have presented a low-cost, hybrid model which is suitable for recognition of low-resource language like handwritten Devanagari characters. The hybrid classifier is a combination of HOG-CNN-MLP algorithms, where HOG is responsible for extracting gradient features from the character images, whereas CNN layers reinforce the particular feature into a specified version of itself. Final classification is done using a multilayer perceptron (MLP). The main contribution of the article lies in the minimal system requirement and overwhelming performance. The model is tested on two public databases of handwritten Devanagari characters DHCD dataset and Kaggle Devanagari offline handwritten character dataset and achieved accuracy of 98.79 and 97.10%, respectively.