Distortion, rotation and scale invariant recognition of hollow Hindi characters

Computer vision is a very vast concept and a lot of researchers are inventing new ideas to improve recognition accuracy for the machine. Today one can see driverless cars, automatic robots doing many human-like activities like playing, dancing, and even in defence service. With the invention of machine learning and deep learning models, the ability of prediction is also improving drastically. The ability of machines to understand printed or handwritten text is also improved these days. Accurate software tools are available for text recognition. The recognition of optical characters is a very mature concept in the Roman script but is in the developing stage for Devanagari script. These OCR systems are producing accurate results in clean printing but perform very poorly when the printing quality is not up to the mark. The performance is even degraded when the characters are distorted or very badly printed/scanned. We have collected 3900 distorted Hindi characters. These characters are hollow in style and randomly rotated and highly distorted. The size of these characters is also varying randomly. The authors have tried to extract six different types of features from these characters to analyze the recognition accuracy and achieved maximum recognition accuracy of 91.1%.