Temporal Convolution based Skip Connections for Online Arabic Handwriting Recognition

Abstract

Handwriting recognition is an active research area in document analysis field. In the few last decades, a rapid growth and a frequent use of data entry devices was observed. Consequently, several approaches have focused on online handwriting modeling and recognition. In this paper, we presented a new system for online Arabic handwriting recognition based on beta-elliptic modeling and one dimensional Residual Neural Networks. Beta-elliptic model was applied to extract the dynamic and geometric characteristics of the trajectory, whereas the developed Residual Network is based on temporal convolution and skip connections and it has ability to represent the sequential aspect of the input data. The experiments have been done on the public Arabic handwriting dataset LMCA and showed the effectiveness of the proposed recognition model that reached the accuracy of 96.87%.