Handwritten Bengali Number Detection using Region Proposal Network

Abstract

As a seventh most spoken native language, Bengali needs a robust and accurate optical character recognition (OCR) especially for number detection. As there is no publicly available well-organized dataset for Bangla number OCR, a synthesized dataset was generated to fill the lack of available data. The recent advancement in artificial intelligence using deep neural networks easily outperforms prior hand selected feature-based machine learning approaches. As the region proposal networks (RPN) in deep neural networks perform very well in detecting objects, it can be used for digit detection in an image. So, in this work a very robust Bengali handwritten number detection system is presented where with the help of deep neural networks and a very well-organized, unbiased generated dataset we achieved state of the art result in handwritten Bangla number detection. This system beats any related prior works by a large margin while considering a real world dataset for benchmarks. The overall detection accuracy was 97.8%. The processing can be done real-time with about 35 images per second using a GPU. Also, while implementing the solution is completely based on python, the framework used for deep learning is Google’s Tensorflow and the dependencies, all of which are publicly available.