Farsi Digit Recognition Using GAN-Generated Data and Convolutional Neural Networks

Abstract

Handwritten digit recognition is one of the most active study areas in computer vision because to its numerous applications such as automatically identifying the digits in bank checks and car numbers. Handwritten Latin digits have been the subject of extensive research over the last three decades, whereas Persian handwritten digits have received far less attention. For this reason, we will concentrate on the problem of recognizing Persian (Farsi) handwritten numerals. The main challenge in the recognition of Persian handwritten digits is the presence of different patterns in Persian digit writing, which complicates the feature extraction process. An appropriate approach for automated feature extraction has been the focus of most earlier investigations since handcrafted feature extraction methods are complex and have unstable performance levels. This paper studies the use of a dataset of Persian handwritten digits generated by a Generative Adversarial Network (GAN) to develop a highly accurate Convolutional Neural Network (CNN) model for digit recognition. The proposed CNN architecture achieved a test accuracy of 99.7%, demonstrating its effectiveness. This study highlights the viability of GAN-generated datasets for machine learning applications, especially in resource-constrained scenarios.