Bangla handwritten character recognition using MobileNet V1 architecture

Tapotosh Ghosh, Md. Min-Ha-Zul Abedin, Shayer Mahmud Chowdhury, Zarin Tasnim, Tajbia Karim, S. M. Salim Reza, Sabrina Saika, Mohammad Abu Yousuf


Handwritten character recognition is a very tough task in case of complex shaped alphabet set like Bangla script. As optical character recognition (OCR) has a huge application in mobile devices, model needs to be suitable for mobile applications. Many researches have been performed in this arena but none of them achieved satisfactory accuracy or could not detect more than 200 characters. MobileNet is a state of art (convolutional neural network) CNN architecture which is designed for mobile devices as it requires less computing power. In this paper, we used MobileNet for handwritten character recognition. It has achieved 96.46% accuracy in recognizing 231 classes (171 compound, 50 basic and 10 numerals), 96.17% accuracy in 171 compound character classes, 98.37% accuracy in 50 basic character classes and 99.56% accuracy in 10 numeral character classes.



Convolutional neural network; Handwritten; Mobile application; MobileNet; Optical character recognition

Full Text:




  • There are currently no refbacks.

Bulletin of EEI Stats