Temporal Residual Network Based Multi-Head Attention Model for Arabic Handwriting Recognition

In this study, we developed a new system for online Arabic handwriting recognition based on temporal residual networks with multi-head attention model. The main idea behind the application of attention mechanism was to focus on the most relevant parts of the data by a weighted combination of all input sequences. Moreover, we applied beta elliptic approach to represent both kinematic and geometric aspects of the handwriting motion. This approach consists of representing the neuromuscular impulses involving during the writing act. In the dynamic profile, the curvilinear velocity can be fitted by an algebraic sum of overlapped beta functions, while the original trajectory can be rebuilt by elliptic arcs delimited between successive extremum velocity instants. The experiments were conducted on LMCA database containing the trajectory coordinates of 23141 Arabic handwriting letters, and showed very promising results that achieved the recognition rate of 97,12%.