Online Arabic Handwriting Recognition Using Graphemes Segmentation and Deep Learning Recurrent Neural Networks

Abstract

Currently, many research initiatives in online handwriting recognition field have been investigated. The challenge is more earnest for Arabic scripts due to some constraints such as the existence of similar shapes of characters and their inherent cursiveness, etc. The present study highlights a new framework for online Arabic handwriting recognition based on graphemes segmentation and two existing models of recurrent neural network (RNN)—long short-term memory (LSTM) and Bidirectional LSTM (BLSTM). After handwriting signal preprocessing, the developed algorithm proceeds by the detection of the script baseline considering the accordance between the alignment of its trajectory points and their tangent directions. Then, the handwritten words or pseudo-words are segmented in continuous part called graphemes delimited by the points of ligature bottoms neighboring the baseline. Next, each grapheme is described by a set of relevant features combining static and dynamic features obtained by employing beta-elliptic model, geometric features that contain Fourier descriptors for trajectory shape modeling and other normalized parameters modeling the grapheme dimensions, positions with respect to the baseline and the assignment diacritics codes. The extracted sequences of features vectors are subsequently used as input for the recognition module employing both LSTM and BLSTM version of RNN. Experimental results using the benchmarking ADAB database of online Arabic handwriting show the performance of the proposed system which exceeds the best result of other advanced states of-the-art works.