PuSH - Publikationsserver des Helmholtz Zentrums München

Li, J.* ; Zanca, D.* ; Christlein, V.* ; Hamann, T.* ; Barth, J.* ; Kämpf, P.* ; Eskofier, B.M.

Enhancing IMU-Based Online Handwriting Recognition via Contrastive Learning with Zero Inference Overhead.

In: (Document Analysis and Recognition – ICDAR 2026). Berlin [u.a.]: Springer, 2026. 695 - 711 (Lect. Notes Comput. Sc. ; 16974 LNCS)
DOI
Online handwriting recognition from inertial measurement units enables handwriting on paper as input for digital devices. Running recognition on edge hardware improves privacy and lowers latency, but entails memory constraints. To address this, we propose Error-Aware Contrastive-Enhanced Handwriting Recognition (ECHWR), a training-time contrastive learning framework designed to improve recognition accuracy without increasing inference costs. During training, ECHWR utilizes a temporary auxiliary branch that aligns IMU signal embeddings with semantic text embeddings in a shared space. This is enforced with a dual objective: (i) an in-batch contrastive loss that aligns IMU and text embeddings, and (ii) a hard-negative contrastive loss that uses synthetically perturbed text targets. The auxiliary branch is discarded after training, which allows the deployed model to keep its original, efficient architecture. Experiments on the OnHW-Words500 dataset show that ECHWR significantly outperforms state-of-the-art baselines, reducing character error rates by up to 7.4% on the writer-independent split and 10.4% on the writer-dependent split. Ablations indicate that the hard-negative objective is particularly effective for generalization to unseen writing styles, making our approach well-suited for online handwriting recognition on unseen users. Code is available at: https://github.com/jindongli24/ECHWR.
Altmetric
Weitere Metriken?
[➜Einloggen]
Zusatzinfos bearbeiten [➜Einloggen]
Publikationstyp Artikel: Konferenzbeitrag
Schlagwörter Handwriting Recognition ; Inference ; Handwriting ; Feature (linguistics) ; Intelligent Character Recognition ; Overhead (engineering) ; Pattern Recognition (psychology) ; Sketch Recognition ; Representation (politics)
ISSN (print) / ISBN 0302-9743
e-ISSN 1611-3349
Konferenztitel Document Analysis and Recognition – ICDAR 2026
Zeitschrift Lecture Notes in Computer Science
Quellenangaben Band: 16974 LNCS, Heft: , Seiten: 695 - 711 Artikelnummer: , Supplement: ,
Verlag Springer
Verlagsort Berlin [u.a.]
Institut(e) Institute of AI for Health (AIH)