Bangla Handwritten Digit Recognition Based on Different Pixel Matrices

Abstract

Handwritten digit or character recognition is the primary task of researching in the field of OCR system. But handwritten digit recognition is a complex task due to the variation of training image size and handwritten style. This paper presents how the nature of the motion of the digit recognition model changes based on different pixel images. We have proposed a 12-layer Convolutional Neural Network (CNN) using four different pixel handwritten images such as 24 × 24, 28 × 28, 32 × 32, and 36 × 36. All fundamental steps such as dataset collection, dataset preparation, detail architecture of CNN, and hyperparameter optimization are described briefly in this paper. The goal of this study is to provide a clear insight into CNN hyperparameters, CNN architecture, images of different pixels on which the effectiveness of the model depends. In this research, we have used two publicly available datasets named CMATERDB 3.1.1 and BanglaLekha-Isolated. The proposed method has achieved a validation accuracy of 99.50% for the CMATERDB 3.1.1 dataset and 98.85% for the BanglaLekha-Isolated dataset.KeywordsBangla handwritten recognitionConvolutional neural networkComputer visionDigit recognition