Advancing Handwritten Urdu Character Recognition: A Deep Learning Approach

Abstract

Optical Character Recognition (OCR) has emerged as a prominent field within Artificial Intelligence (AI) and is extensively researched in the domain of pattern recognition (PR). In recent times, OCR has garnered significant attention due to its pivotal role in facilitating the computer's ability to identify and interpret script present in images and documents. A wide variety of scripts, each requiring digitization and recognition, adds to the complexity of the OCR task. In this research, we present the development of a deep learning based model specifically tailored for recognizing handwritten isolated Urdu characters. Our model employs Convolutional Neural Networks (CNNs) for efficient feature extraction from the input images. To evaluate the model's performance, we utilized the UHAT dataset, which consists of 28,328 training images and 4,880 testing images. The CNN model achieved an impressive recognition rate of 99.39% over 100 training epochs on the UHAT dataset. Furthermore, we curated a custom dataset, categorizing it into distinct training and testing subsets. The custom dataset encompasses 6,300 images, partitioned into an 80% training set and a 20% testing set. Our proposed model underwent training on 80% of the custom dataset and achieved a commendable recognition rate of 99.19% on the handwritten character images during testing. The results of this study demonstrate the effectiveness of our deep learning-based approach for Urdu character recognition, paving the way for enhanced OCR capabilities in handling diverse scripts and contributing to the advancement of pattern recognition technologies.