In recent years, the use of Deep Learning has achieved remarkable and great results in various fields, particularly in the domain of English character recognition, whether in written or printed form. This is because English is widely spoken as a first or second language by a significant portion of the world's population. Therefore, Arab researchers have turned to applying the advancements in modern English language processing and similar languages like French to the Arabic language and its equivalents, such as Urdu and Pashto, in order to keep up with global changes. Dealing with the Arabic language and its similar languages like Urdu and Pashto using machine learning and deep learning techniques presents challenges that researchers are trying to overcome to obtain satisfactory and comparable results to those achieved in the English language. Therefore, in this research, the most popular methods, techniques, and technologies in deep learning, such as CNN, LSTM, Bi-LSTM, GRU, and Bi-GRU, were applied and tested on three Arabic language datasets: AHCD, Hijja, and AHDD. Subsequently, a comparison was made between the results obtained from these different techniques using performance measurement methods such as Precision, Recall, and Accuracy.

Recognizing handwritten Arabic characters poses a significant challenge due to the complexities of the cursive script and the visual similarities between characters. While deep learning techniques have shown substantial promise, advancements in model architectures are essential to further enhance performance. Neutrosophic Sets (NS) have demonstrated their potential in improving classification models by effectively handling indeterminate and inconsistent data. This paper introduces a novel approach that integrates Neutrosophic Sets with a hybrid deep learning model, combining Convolutional Neural Networks (CNNs) with Bidirectional Recurrent Neural Networks (Bi-LSTM and Bi-GRU). This integration allows for the extraction of spatial features and modeling of temporal dynamics in handwritten Arabic text. Experiments conducted on the Hijjaa and AHCD datasets revealed that the NS_CNN_Bi-LSTM model achieved an accuracy of 92.38% on the Hijjaa dataset, while the NS_CNN_Bi-GRU model attained 97.38% accuracy on the AHCD dataset, outperforming previous deep learning approaches. These results highlight the significant performance improvements achieved through advanced temporal modeling and contextual representation, without the need for explicit segmentation. The findings contribute to the ongoing development of highly accurate and sophisticated deep learning systems for Arabic handwriting recognition, with broad applications in areas requiring efficient extraction of text from handwritten documents.