Revitalizing Arabic Character Classification: Unleashing the Power of Deep Learning with Transfer Learning and Data Augmentation Techniques

• Published in: Arabian journal for science and engineering (2011) Vol. 49; no. 9; pp. 12791 - 12815
• Main Authors: Amara, Marwa, Smairi, Nadia, Mnasri, Sami, Zidouri, Abdelmalek
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 2024; Springer Nature B.V
• Subjects: Accuracy; Character recognition; Classification; Data augmentation; Datasets; Deep learning; Engineering; Handwriting recognition; Humanities and Social Sciences; Knowledge management; Machine learning; multidisciplinary; Performance enhancement; Research Article-Computer Engineering and Computer Science; Science; Deep learning; Data augmentation; CNN; Handwritten digit recognition; Transfer learning; Arabic character
• ISSN: 2193-567X; 1319-8025; 2191-4281
• DOI: 10.1007/s13369-024-08818-9

Deep learning techniques have demonstrated remarkable success in various domains, including character classification tasks. However, the performance of deep learning models heavily relies on the availability of large-annotated datasets. This research work is motivated by the need to overcome the difficulties associated with handwritten Arabic character recognition, and the constraints provided by limited training data. It is also motivated by the need to enhance the model’s generalizability to unknown characteristics and to improve the accuracy of deep learning character classification models. To overcome this limitation, we apply transfer learning and data augmentation strategies to improve the character classification using deep learning. The proposed model transfers knowledge from previously trained models via transfer learning, addresses data scarcity, and reflects generalizable properties. Indeed, we utilize a VGG16-ImageNet transfer learning model, which is systematically enhanced through data augmentation across three distinct models: pre-trained ImageNet weights with a frozen backbone, pre-trained ImageNet weights with a fine-tuned backbone, and initiating with a randomly initialized. In each case, data augmentation plays a critical role. Our experimental results show that better precision and recall values were recorded for most classes in our dataset, which indicates the model’s ability to accurately identify instances of each character. Moreover, when applying our method to the IFHCDB and HACDB datasets, we observed an impressive recognition accuracy of 96.01% and 97.15%, respectively. This clearly indicates that involving transfer learning and data augmentation significantly improves the performance of deep learning models, especially for small size training datasets.