Online Handwriting Analysis and Recognition using Beta-elliptic model and deep BLSTM Network

Résumé

This thesis highlights the progress made in the fields of online handwritten recognition and analysis, as well as associated applications utilizing artificial intelligence techniques. Indeed, automatic analysis and recognition of handwriting are complex tasks that combine various strategies and methods developed around computer vision. Initially, we described the different families of systems used for handwriting analysis and recognition, which allowed us to outline the state of the art in this field. This study subsequently enabled us to develop a handwriting recognition system for letters and words to address various challenges. In our work, we applied two methods for modeling online handwriting: the beta-elliptic model and grapheme-based segmentation. Specifically, we exploit the dynamic and geometric characteristics of the beta-elliptic model for modeling the trajectory of online handwriting. We also leverage the developed trajectory segmentation model, which detects the script’s baseline by aligning trajectory points with their tangent directions. In this latter model, handwritten words or pseudo-words are segmented into continuous parts called graphemes, which are delimited by ligature valleys neighboring the baseline. The first objective of our thesis project is to simulate the beta-elliptic approach using a deep stacked recurrent neural network of the BLSTM (Bidirectional Long Short-Term Memory) type. The proposed architecture encompasses preprocessing, segmentation, and trajectory approximation through two profiles—dynamic (velocity) and geometric—using neural computation sequences. We used and compared various proposed models for online character and word recognition in a predefined context based on LSTM and BLSTM versions of recurrent neural networks. Additionally, we introduced multiple data augmentation methods (geometric method, frequency-based method, and beta-elliptic model-based method) to improve character/word recognition rates within the proposed architectures. In the second part of this thesis, we contributed to the development of an online Arabic handwriting recognition system with an open vocabulary. Research in this area is still in its early stages. In this section, we focused on the proposed neural beta-elliptic model and the grapheme-based segmentation model for feature extraction, along with the deep BLSTM recurrent network followed by the Connectionist Temporal Classification (CTC) algorithm for sequence labeling. Beyond the recognition system, we also developed a mobile application for analyzing and evaluating the quality of online handwriting in multiple languages (Arabic, Urdu, Persian, and Latin) designed for primary school children. This application assesses handwriting quality and provides informative feedback. It is based on multiple criteria, including shape, writing order, kinematic direction, and adherence to the baseline, as well as a combination of different features derived from the neural beta-elliptic model and perceptual convolutional neural networks (CNNs). Specifically, we used beta-elliptic parameters to capture both the visual and dynamic characteristics of beta-segmented strokes forming the handwriting trajectory. Additionally, the CNN model is employed to express the trajectory and its final visual representation after drawing. Finally, this work has demonstrated that the proposed architectures are valid and serve as standard candidate approaches for addressing various computer vision tasks. It paves the way for new developments in handwriting recognition, with further significant improvements expected through the application of other deep learning methods

Citer ce document

Accès au document

Ce document est hébergé sur son dépôt institutionnel d'origine.

Auteur(s)

Statistiques

Consultations : 4

Téléchargements : 0