Defended on November 22th 2024
Design of a deep neural network architecture dedicated to handwriting synthesis from kinematic sensors of a digital pen
This thesis focuses on a digital pen equipped with kinematic sensors, and its aim is to reconstruct the in-line trace of handwriting. We introduce a new processing pipeline that associates pen sensor signals with the corresponding writing trajectory. Based on Dynamic Time Warping to align the signals and an architecture inspired by Temporal Convolutional Networks Additionally, we present a Mixture-Of-Experts (MOE) approach to enhance the focus and understanding of each aspect of handwriting, comprising a touching expert model for pencil touches and a pen-up expert model for pen trajectories. A significant challenge is the variation in captured signals between adults and children, due to differences in speed and confidence in handwriting gestures. We address this through a domain adaptation approach. Furthermore, we introduce a new public benchmark dataset to support future research and comparisons in the field of handwriting reconstruction.
Composition du jury
- Rapporteur – Andreas FISCHER, University of Applied Sciences and Arts Western Switzerland (HES-SO)
- Rapporteur – Clément CHATELAIN, INSA Rouen Normandie
- Examinateur : Nicolas RAGOT, Polytech Tours
- Examinateur : Elisa FROMONT, Université de Rennes
- Directeur de thèse : Eric ANQUETIL, INSA de Rennes
- Co-Directeur de thèse : ROMAIN TAVENARD, université Rennes 2
- Co-encadrant de thèse (invité) : Yann SOULLARD, université Rennes 2
