Handwritten Arabic characters recognition using Capsule Networks
Daldali Mehdi, Abdelghani Souhar · 2022
Handwritten text exhibits a diversity in styles, and unpredictable characteristics making of Handwriting Recognition (HWR) an interesting computer vision problem. In the case of Arabic script, the recognition task is especially more complex, due to its cursive nature, and emphasis on fluid and connected pen strokes, rarely seen in other scripts. Unfortunately current end-to-end approaches fail at modeling the structural aspect of visual information with high accuracy, which is detrimental for tasks such as Optical Character Recognition (OCR). Capsule networks are a neural network architecture, using a set of interesting concepts which can accurately model the structural aspect of image features, to solve some of the Arabic script recognition problems faced by other systems. Our proposed approach based on Capsules yielded interesting results using a data set of around 50 thousand Arabic characters with neither binarization nor noise reduction, while achieving more than 97% TOP-1 accuracy.