ISSN(Online): 2320-9801 ISSN (Print): 2320-9798 International Journal of Innovative Research in Computer and Communication Engineering (An ISO 3297: 2007 Certified Organization) Vol. 4, Issue 2, February 2016 Copyright to IJIRCCE DOI: 10.15680/IJIRCCE.2016. 0402001 1101 Handwritten Digit Recognition Using Convolutional Neural Networks Haider A. Alwzwazy 1 , Hayder M. Albehadili 2 , Younes S. Alwan 3 , Naz E. Islam 4 M.E Student, Dept. of Electrical and Computer Eng. University of Missouri-Columbia, MO, USA 1,2,3 Professor, Dept. of Electrical and Computer Eng. University of Missouri-Columbia, MO, USA 4 ABSTRACT: Recently handwritten digit recognition becomes vital scope and it is appealing many researchers because of its using in variety of machine learning and computer vision applications. However, there are deficient works accomplished on Arabic pattern digits because Arabic digits are more challenging than English patterns. Hence, the lacking research of using Arabic digits endeavours us to dig deeper by creating our challenge Arabic Handwritten Digits which consists of more than 45,000 samples. As a challenging dataset is used for evaluation, a robust deep convolutional neural network is used for classification and superior results are achieved. KEYWORDSHandwritten Digit Recognition; Arabic Handwritten Digits; I. INTRODUCTION Recently Deep Convolutional Neural Networks (CNNs) becomes one of the most appealing approaches and has been a crucial factor in variety of recent success and challenging machine learning applications such as challenge ImageNet [1, 2, 3, 4, 5,24], object detection [1, 6, 7], image segmentation [9,10], and face recognition [11, 12, 13]. Therefore, CNNs is considered our main model for our challenging tasks of image classification. Specifically, it is used for handwriting digits recognition which is one of high academic and business transactions [14]. Handwriting digit recognition application is used in different tasks of our real life time purposes. Precisely, it is used in banks for reading checks, post offices for sorting letter, and many other related tasks. Apparently English Handwriting datasets are widely available and significant achievements have been made for English digit datasets such as CENPARMI [15], CEDAR[16], and MNIST[17], However, there are rare works accomplished on Arabic digit datasets for many reasons. One of critical factor that can influence working on Arabic dataset is lacking to dataset. The unavailability of dataset can be one of the essential factors that can diminish working on Arabic datasets. Hence, deficiency of large challenging Arabic dataset strives us to extensively working on creating a largest and most challenging dataset which contains more than 45,000 patterns. Furthermore, we investigate and demonstrate a powerful DCNN used for classification. Not only designing powerful DCNN is presented but also critical parameters of CNN is carefully selected and tuned to produce final concrete model which achieves superior results. II. RELATED WORK Handwritten digit recognition (HDR) is considered one of trivial and critical machine learning problems. It has been used widely by researchers as experiments for theories of machine learning algorithms for many years. In recent years, neural networks and conventional neural network currently provide the best solutions to many problems in handwritten digit recognition. A novel hybrid CNN–SVM model for handwritten digit recognition is designed by [18]. This hybrid model automatically extracts features from the raw images and generates the predictions. For this work, the author used non-saturating neurons and a very efficient GPU implementation of the convolution operation to reduce overfitting in the fully-connected layers. To enhance method proposed in [8], [19] tackled critical investigations to diminish limitation inherited from [8]. The author introduces a novel visualization technique that gives insight into the function of feature layers and the procedure of the classifier [20] have observed convolutional net architecture that can be used even when the amount of learning data is limited. [21] have used new network structure, called Spatial Pyramid Pooling SPP-net, can generate a fixed-length representation regardless of image. Multi-column DNN (MCDNN) used MNIST digits. The result has a very low 0.23% error rate [22]. Hayder M. Albeahdili et al. [9] have performed a new