Advanced Deep Learning Techniques for Arabic Handwritten Digit Classification

This paper explores the application of deep learning algorithms, particularly Convolutional Neural Networks (CNNs), for Arabic handwritten digit recognition. The dataset used in this study is the Arabic Handwritten Digits Database (AHDD1), comprising 70,000 handwritten Arabic digit images, of which 60,000 were used for training and 10,000 for testing. The dataset covers digits from 0 to 9 and includes 10 samples per digit, provided by 700 contributors from diverse backgrounds, ensuring rich variability in handwriting styles. Our results indicate that deep learning models, especially CNNs, deliver strong performance in recognizing Arabic handwritten digits, achieving high accuracy. Among the tested models, simpler architectures like Multi-Layer Perceptron (MLP) and Fully Connected Neural Networks (FCNN) performed well with lower computational costs. However, the best-performing model in this study was the CNN with Data Augmentation and Dropout, which achieved the highest accuracy of 99.32%. This model demonstrates the effectiveness of advanced techniques, such as data augmentation, in enhancing the model's robustness and performance. While more complex models like ResNet provided slightly lower accuracy, they required significantly longer training times and higher computational resources. The findings underscore the importance of balancing model accuracy with computational efficiency in real-world applications. The study suggests that less complex models, enhanced with dropout regularization and batch normalization, can effectively solve Arabic handwritten digit recognition without excessive computational overhead.