Farsi Handwriting Digit Recognition Based on Convolutional Neural Networks

Abstract

In this paper, a convolutional neural network (CNN) is exploited for Farsi handwritten digit recognition. For training and evaluating the CNN, the "HODA" dataset was used which consists of 80000 images of Farsi handwritten digits. In the proposed method, we focused on the efficient and unique feature of Farsi digits that is using just the half upper part of the digits for recognition purpose. The proposed method, despite of a 50% reduction in the data size which fed to the CNN, yielded an acceptable reduction in time consuming for training and evaluate CNN of about 50 % compared when using the full image of the digits (full data), and just a 1.5% increase in recognition error.