Improving Persian Handwritten Digit Recognition using Convolutional Neural Network

Abstract

Persian digit recognition plays a crucial role in computer vision and pattern recognition. Existing algorithms fall into two categories: traditional methods and deep learning approaches. While many deep learning techniques are documented, they often depend on pre-trained networks with numerous parameters, requiring substantial resources and time for training and prediction. This paper presents a novel convolutional neural network (CNN) architecture for Persian digit recognition that is shallower than current models, thereby reducing the number of trainable parameters. We introduce dilated convolution layers to capture larger features without increasing parameters and propose a combined loss function to improve accuracy. Trained on the HODA dataset, our method achieves a validation accuracy of 99.82\%, test accuracy of 99.79\%, and training accuracy of 100\%. The proposed network demonstrates enhanced accuracy, faster performance, and significantly reduced implementation time due to its streamlined architecture.