Professional Documents
Culture Documents
Abstract—Bangla handwritten digit recognition is a convenient This paper is organized as follows: The related works are
starting point for building an OCR in the Bengali language. presented in Section II. Our dataset and proposed approach
Lack of large and unbiased dataset, Bangla digit recognition was are described in Section III. Section IV shows the
not standardized previously. But in this paper, a large and preprocessing technique of images. We have shown the
unbiased dataset known as NumtaDB is used for Bangla digit architecture of the deep convolutional neural network in
recognition. The challenges of the NumtaDB dataset are highly Section V. Section VI represents our experiments and result
unprocessed and augmented images. So different kinds of analysis and finally conclusion and future works are briefed in
preprocessing techniques are used for processing images and Section VII.
deep convolutional neural network (CNN) is used as the
classification model in this paper. The deep convolutional II. RELATED WORKS
neural network model has shown an excellent performance,
securing the 13th position with 92.72% testing accuracy in the There are several research works based on Bangla
Bengali handwritten digit recognition challenge 2018 among 57 handwritten digit recognition using deep learning. But most of
participating teams. A study of the network performance on the the research works used biased dataset like CMATERDB
MNIST and EMNIST datasets were performed in order to 3.1.1 [4] because NumtaDB dataset was not available then.
bolster the analysis. Recently some researchers have shown a better accuracy of
99.50% using the auto encoder and deep convolutional neural
Keywords— NumtaDB, CNN, Bangla Handwritten Digit network for CMATERDB 3.1.1 dataset. The local binary
Recognition, Large Unbiased Dataset, Computer Vision pattern was also used for Bangla digits recognition [5]. The
first CNN architecture known as Le-net was also used by some
I. INTRODUCTION researchers for Bangla digit recognition [6]. The
Bangla handwritten digit recognition is a classical Convolutional neural network was introduced for better
problem in the field of computer vision. There are various supervised learning and accuracy [7]. So recent researchers
kinds of practical application of this system such as OCR, of Bangla handwritten digit recognition are using deep CNN
postal code recognition, license plate recognition, bank checks architecture. There are some other classifiers like Support
recognition etc. [1]. Recognizing Bangla digit from Vector Machine (SVM), Neural Network (NN) etc. for
documents is becoming more important [2]. The unique handwritten digit recognition. But the performance of CNN is
number of Bangla digits are total 10. So the recognition task better than other classifiers. As a result, CNN has become the
is to classify 10 different classes. The critical task of recent trend handwritten digit recognition. Modified CNN
handwritten digit recognition is recognizing unique architecture by adding more layers or more nodes has become
handwritten digits. Because every human has his own writing a way to break the state of the art accuracy. This strategy is
styles. But our contribution is for the more challenging task. followed by some research works [8]. Besides Bangla, there
The challenging task is about getting robust performance and are some significant research works in English handwritten
high accuracy for large, unbiased, unprocessed, and highly digit recognition. MNIST [9] and EMNIST [10] are the most
augmented NumtaDB [3] dataset. The dataset is a popular dataset for English handwritten digit and character
combination of six datasets that were gathered from different recognition. EMNIST [10] presented handwritten digit
sources and at different times containing blurring, noise, recognition for both balanced and imbalanced dataset. We
rotation, translation, shear, zooming, height/width shift, researched considering MNIST and EMNIST dataset
brightness, contrast, occlusions, and superimposition. We previously. The research work of this paper is related to our
have not processed all kinds of augmentation of this dataset. previous research and that was English handwritten digit
We have processed blur and noisy images mainly. Then our recognition. We achieved better results from our previous
processed images are classified by a deep convolutional neural research than the previous research [10] and it motivates us to
network. contribute in Bangla handwritten digit recognition too.
B. Proposed Approach
We propose to classify NumtaDB handwritten digits
following two major steps. These steps are:
• Preprocessing of images.
• Deep convolutional neural network.