HTR-ViTRNN: A CTC-Aware Temporal Consistency Approach for Resource-Constrained Handwritten Text Recognition
IEEE Access · 2026 · 14 · 115825-115842
DOI: 10.1109/access.2026.3715574 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Pham Doan Tinh und 1 weiteren Autor:innen. Die Arbeit nennt 45 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Handwritten Text Recognition (HTR) models often exhibit optimization difficulty when trained with small batch sizes. In this study, we investigate the temporal behavior of CTC logits in Connectionist Temporal Classification (CTC)-based HTR models, especially under limited training conditions that require a small batch size. Vision Transformer-based architectures such as HTR-VT have shown competitive performance; however, their feed-forward structure does not explicitly impose local temporal constraints on adjacent output steps, leading to ambiguous or fragmented CTC alignments under small-batch training. We propose HTR-ViTRNN, a lightweight logit-level temporal regularization approach that integrates a Recurrent Neural Network (RNN) after the HTR-VT visual encoder to improve CTC alignment consistency under small-batch CTC training. The RNN component enables modeling sequential relationships that the encoder does not fully capture, encouraging temporal consistency in CTC logits without changing CTC decoding; the default one-layer BiLSTM with 256 hidden units per direction increases the parameter count from 53.5M to 56.0M. Experiments on the IAM, LAM, and READ 2016 datasets demonstrate relative CER and WER reductions of up to 20.5% and 17.5%, respectively, across batch sizes (32, 16, 8, 6, and 4) compared to HTR-VT, together with improved alignment consistency and convergence behavior under small-batch training. Notably, to achieve comparable accuracy on the test datasets, the proposed model can operate with smaller batch sizes, resulting in shorter training time compared with HTR-VT. These results suggest that recurrent logit regularization can improve CTC alignment consistency under constrained training regimes. The source code is publicly available at: https://github.com/phamdoantinh/HTR-ViTRNN.
Zitieren
APA
Tinh, P. D., & An, H. H. (2026). HTR-ViTRNN: A CTC-Aware Temporal Consistency Approach for Resource-Constrained Handwritten Text Recognition. IEEE Access, 14, 115825-115842. https://doi.org/10.1109/access.2026.3715574
Vancouver
Tinh PD, An HH. HTR-ViTRNN: A CTC-Aware Temporal Consistency Approach for Resource-Constrained Handwritten Text Recognition. IEEE Access. 2026;14:115825-115842. doi:10.1109/access.2026.3715574
BibTeX
@article{Tinh2026,
author = {Tinh, Pham Doan and An, Ha Huu},
title = {HTR-ViTRNN: A CTC-Aware Temporal Consistency Approach for Resource-Constrained Handwritten Text Recognition},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {115825--115842},
doi = {10.1109/access.2026.3715574}
}
EUVIMED