Combining handwriting dynamics with the modern ConvNeXtV2 convolutional backbone for handwritten character recognition on Russian- and English-language Datasets

Combining handwriting dynamics with the modern ConvNeXtV2 convolutional backbone for handwritten character recognition on Russian- and English-language Datasets

Iatsenko D.V. (IHTPT SFedU, Rostov-on-Don, Russia)

Abstract

This work extends the author’s previous research on leveraging dynamic pen-motion characteristics to improve handwritten text recognition. We combine the earlier proposed dynamic component–angular and spectral trajectory features–with a modern visual backbone, ConvNeXtV2_tiny, and evaluate on three datasets: EMNIST (by_class), UJI Pen Characters 2, and our own Russian Handwritings Tracked. We show that, after moderate augmentation tuning, the visual branch achieves state-of-the-art performance on EMNIST among the models compared, while the dynamic branch and their ensemble improve robustness on datasets with greater handwriting variability (UJI, Russian Handwritings Tracked). The results confirm the complementarity of visual and kinematic features and highlight the promise of the method for “difficult handwriting” scenarios.

Keywords

handwriting recognition; online handwriting; kinematic features; pen trajectory; spectral features; ConvNeXtV2; EMNIST; UJI2; Russian Handwritings Tracked; ensembling.

Edition

Proceedings of the Institute for System Programming, vol. 38, issue 2, 2026, pp. 195-206

ISSN 2220-6426 (Online), ISSN 2079-8156 (Print).

DOI: 10.15514/ISPRAS-2026-38(2)-13

For citation

Iatsenko D.V. Combining handwriting dynamics with the modern ConvNeXtV2 convolutional backbone for handwritten character recognition on Russian- and English-language Datasets. Proceedings of the Institute for System Programming, vol. 38, issue 2, 2026, pp. 195-206 DOI: 10.15514/ISPRAS-2026-38(2)-13.

Full text of the paper in pdf (in Russian) Back to the contents of the volume