Abstract:
Boosted performance in Offline Arabic Handwritten Recognition (OAHR) models have recently gained significance due to the variety of the growth of the real-world applications in the pattern recognition domain. However, OAHR systems have been suffering from numerous challenges related either to the Arabic script intrinsic characteristics (e.g., cursive nature, overlapping, ligatures, dots) or to the Handwriting Recognition (HWR) systems (i.e., unlimited human handwriting styles; databases mostly lacking availability and sufficiency; having no unified, well-performed, and automatic end-to-end model that is generalized enough to recognize Arabic digits, characters, and words).
The Deep Learning (DL) and Deep Transfer Learning (DTL) techniques have acquired a remarkable performance in the field of text classification; significantly, the Convolutional Neural Network (CNN) has achieved superior performance compared to other regular Neural Networks (NN) by learning the intricate and high-dimensional features and perform the classification task automatically into one end-to-end process.
The developed Deep CNN (DCNN) architecturesusing incremental hand-crafted design methodology contributed with four types of novel models encompassing a DCNN model, an enhanced very DCNN model named VDCNN, four pre-trained models, and a VDCNN model based on TL techniques named VDCNN-based DTL. The evaluated DCNN models haveachieved outstanding recognition accuracy compared to the existing state-of-the-art OAHR DL-based approaches that have been trained on an assorted set of six benchmark datasets, including MADBase (Digits), SUST-ALT (Digits), CMATERDB (Digits), SUST-ALT (Characters), HACDB (Characters), and SUST-ALT (Names). DCNN model achieved recognition accuracies of 99.91%, 99.82%, 99.72%, 99.86%, 99.91%, 99.95%, in order, while the VDCNN model achieved recognition accuracies of 99.89%, 99.87%, 99.78%, 99.89%, 99.93%, 99.95%, respectively. The VDCNN-based DTL model has transcended DCNN and VDCNN models’ performance with significant minimal training time and testing accuracy of 99.90%, 99.94% using the selected CMATERDB and HACDB databases, respectively.
A further experimental study is conducted on the used benchmark Arabic databases by exploiting DTL-based feature extraction, demonstrating the DCNN model's superiority in relation to state-of-the-art VGGNet-19 and MobileNet pre-trained models.
Finally, DCNN and VDCNN models have an outstanding potential for recognizing other languages, such as the MNIST English digits database, in which they obtained an accuracy of 99.94%, 99.90%, respectively.
