Deep learned BLSTM for online handwriting modeling simulating the Beta-Elliptic approach

• Published in: Engineering science and technology, an international journal Vol. 35; p. 101215
• Main Authors: Hamdi, Yahia, Boubaker, Houcine, Rabhi, Besma, Qahtani, Abdulrahman M., Alharithi, Fahd S., Almutiry, Omar, Dhahri, Habib, Alimi, Adel M.
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.11.2022; Elsevier
• Subjects: Beta-elliptic model; BLSTM; Deep Learning; Online Handwriting modeling; RNN; Beta-elliptic model; Online Handwriting modeling; RNN; Deep Learning; BLSTM
• ISSN: 2215-0986; 2215-0986
• DOI: 10.1016/j.jestch.2022.101215

[Display omitted] The beta-elliptic model (BEM) has proven a great success in several applications such as handwriting recognition and analysis, handwriting identification, the effect of age on the kinematics of hand movements, etc. With the emergence of deep learning technologies during the last years and their application in several fields, the implementation of beta-elliptic model with a simple multi-stage deep learned recurrent neural network (RNN) is required. In this paper, we propose a new online handwriting trajectory modeling by simulating the beta-elliptical approach to limit the calculation time and to have an end-to-end description system meeting the needs of mobile device users. The developed model deploys a multi-stage architecture based deep learned recurrent neural network (RNN) with bidirectional Long Short-Term Memory (BLSTM) that simulates the process of extracting the dynamic and geometric parameters composing the beta-elliptical vector. This architecture encompasses the pre-processing, segmentation, and approximation steps of the trajectory in its two velocity and geometric profiles and is traced by only neural computation sequences. To evaluate our model, a similarity degree between the beta-elliptic and BLSTM estimation models is measured by means of MAE (mean absolute error) and RMSE (Root Mean Square Error) metrics. Experimental results on LMCA and ADAB datasets show the efficiency of the proposed RNN model for online handwriting trajectory modeling consisting of 3.75%, 5.26% for RMSE, and 1.69%, 2.75% for MAE respectively.