Automatic arabic handwritten characters Recognition using ensemble of convolutional neural networks from scratch

Arabic handwritten character recognition (AHCR) presents significant challenges owing to the diversity and complexity of the script, compounded by variations in handwriting styles, sizes, and orientations. Using an ensemble of convolutional neural networks (CNNs), this study aims to develop a robust and accurate recognition system. To extract features from the input images effectively, we designed three distinct CNN models, each comprising unique combinations of network layers and parameters. Advanced normalization techniques have been incorporated to handle variations in handwriting styles. In addition to the standard ensemble, to further improve performance we experimented with a weighted average ensemble. Also, to identify the optimal weights, hyperparameter tuning was conducted using a grid search. This process resulted in a maximum classification accuracy of 98.0654%, with a precision of 97.36%, a recall of 97.21%, and an F1-score using the weighted average ensemble model of the three CNNs from scratch. The model achieves the lowest average misclassification rate (1.934%). By effectively addressing the challenges posed by handwriting variations and demonstrating substantial improvements in recognition accuracy, this approach shows potential for practical application in automated AHCR systems.