Deep Learning Optimization Methods for Arabic Handwritten Recognition Using CNN

Abstract Deep learning has achieved great success in image recognition tasks. Arabic is the sixth most spoken language globally, with 274 million speakers 1, making handwritten Arabic recognition a significant challenge due to variations in writing styles and connectivity of characters. This paper presents a deep learning approach for Arabic handwritten digit and character recognition using Convolutional Neural Networks (CNNs). We implement a 4-layer CNN architecture with batch normalization, dropout, and multiple optimizers (Adam, RMSprop, Adagrad, Nadam) to evaluate performance on two datasets: MADBase (60,000 digit images) and an Arabic character dataset (16,800 images). Our model achieves 98.86% validation accuracy with Adam optimization, uniform initialization, and ReLU activation, significantly outperforming baseline methods (32.37% accuracy). The system demonstrates robustness through extended training (20 epochs) while maintaining stability, with check pointing preserving optimal weights. This work contributes to improving Arabic handwriting recognition for applications such as document digitization, archival preservation, and automated data entry.