.png){kind=icon}
.png){kind=icon}
Return
TlMamba: structure-aware Vision Mamba for long-tailed ancient Tai Lue palm-leaf manuscript character recognition
.png){kind=icon}
PRE
AI
J
Z
Z
Y
DOI：10.1007/s00371-026-04680-y.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Recognizing handwritten characters in ancient Tai Lue palm-leaf manuscripts poses significant challenges due to high inter-character similarity, visually ambiguous glyphs, and long-tailed category distributions. This paper introduces TlMamba, a structure-aware Vision Mamba framework for ancient Tai Lue palm-leaf manuscript character recognition. TlMamba strengthens fine-grained local feature modeling through structure-aware patch embedding, integrates local stroke structures and contextual relationships through a structure-semantics dual-branch design, and uses two-stage knowledge distillation to improve rare-character learning under limited-sample conditions. We construct HLDLC1.0 as a character-level dataset containing 3739 samples from 51 ancient Tai Lue character categories. Experimental results on HLDLC1.0 show that TlMamba achieves 96.17% overall accuracy and 95.28% Tail(25%) accuracy, indicating a relatively high level of overall recognition performance on this dataset while maintaining favorable performance in tail-category recognition. Cross-dataset evaluations on Oracle-MNIST, Devanagari, and Tamil datasets and additional manuscript-related datasets further support its applicability to related character-level recognition tasks. Source code, trained model weights, fixed train/validation/test splits, configuration files, environment instructions, HLDLC1.0 documentation, and related scripts are publicly available at https://github.com/shangzhengshuo123/TlMamba to support reproducible research in ancient manuscript character recognition.
Keywords:
Ancient manuscript recognition
Tai Lue palm-leaf manuscripts
Handwritten character recognition
Vision Mamba
State space models
Structure-aware representation learning
Long-tailed recognition
Low-resource OCR
Cultural heritage digitization
Fine-grained visual classification
Journal
IF:
2.9
Papers:
4.5K
Citations:
6.5K
Organization
Cited Papers