Arabic Handwritten Digits Dataset

Abstract:

Handwritten digit recognition is an essential area of research within computer vision and deep learning due to its broad range of practical applications. Recognizing Arabic handwritten digits presents unique challenges compared to Latin-based scripts, primarily due to the variation in handwriting styles, cultural influences, and the intricacies of Arabic calligraphy.

The dataset was developed with these challenges in mind and is suitable for training and evaluating deep learning models aimed at recognizing handwritten Arabic digits. Leveraging deep learning techniques, particularly Convolutional Neural Networks (CNNs), has shown significant improvement in recognition accuracy over traditional machine learning methods. Models like LeNet-5 have been effectively utilized to train on this dataset, achieving an impressive accuracy of 99.15%.

Key Features:

• Data Size: 70,000 images (60,000 for training, 10,000 for testing).
• Writers: 700 different individuals.
• Digit Range: The dataset contains Arabic digits ranging from 0 to 9.
• Image Format: Grayscale, 28×28 pixels.
• Diversity: The dataset encompasses a wide range of handwriting styles, collected from individuals of different ages and backgrounds, including students from engineering, law, and medical schools, as well as governmental employees.

Context and Significance:

The field of Arabic handwritten digit recognition has grown substantially, driven by the need for automation in various sectors such as education, finance, and public services. However, due to the complex nature of Arabic script and its handwritten form, recognizing Arabic digits accurately requires sophisticated models trained on large and diverse datasets.