HYBRID DEEP LEARNING-CROW SEARCH ALGORITHM FOR ACCURATE ARABIC HANDWRITTEN DIGIT RECOGNITION

Abstract

This paper proposes a hybrid deep learning framework for Arabic handwritten digit recognition by optimizing Convolutional Neural Network (CNN) hyperparameters using the Crow Search Algorithm (CSA). Due to the high variability and structural complexity of Arabic handwritten digits, achieving optimal CNN performance requires efficient and automatic hyperparameter tuning. In the proposed approach, CSA is employed to optimize key CNN hyperparameters, including filter size, number of filters, mini-batch size, and learning rate, with the objective of minimizing classification error. The MADBase dataset is used for evaluation, and preprocessing steps such as normalization, image reshaping, one-hot encoding, noise reduction, and data shuffling are applied to enhance training efficiency and model robustness. The CNN architecture is trained using the optimized hyperparameters obtained through CSA iterations. Experimental results show that the proposed CSA-optimized CNN achieves 99% accuracy on the testing set, demonstrating strong generalization capability and stability. The findings confirm that CSA effectively improves CNN performance while eliminating the need for manual hyperparameter tuning, making the framework suitable for other image classification tasks.