The Challenge of Glyph Classification
The digitization of ancient palm leaf manuscripts is critical for cultural preservation, but it presents significant technical hurdles. Isolated glyph (character) classification is a foundational task, yet it is complicated by several factors: the degradation of the physical manuscripts, complex character classes with subtle variations, and the immense visual similarity between different glyphs. Furthermore, the scarcity of large, labeled datasets for these low-resource languages makes it difficult to train robust deep learning models.
Our research focuses on overcoming these obstacles by developing novel model architectures and efficient data-handling strategies. The following publications detail our core contributions in this area.
Key Research & Publications
Improving Isolated Glyph Classification Task for Palm Leaf Manuscripts
N. Thuon, J. Du, J. Zhang. ICFHR 2022.
Summary: This foundational work investigates methods to enhance glyph classification by focusing on both front-end (data-centric) and back-end (model-centric) processes. We explore multi-task pre-processing, including data augmentation and image enhancement, to improve dataset quality. We then conduct a comprehensive analysis of various deep learning backbones, comparing the performance of established CNNs (VGG, ResNet, EfficientNet) against modern attention-based models (ViT, DeiT, CvT). The experiments, conducted on Balinese, Sundanese, and Khmer scripts, provide a clear benchmark and demonstrate effective training strategies for the document analysis community.
KhmerFormer: Multi-Scale CNNs-Transformer with External Attention for Ancient Khmer Isolated Glyph Classification
N. Thuon, J. Du, APSIPA ASC 2024.
Summary: To address the high intra-class variations and inter-class similarities in Khmer glyphs, this paper introduces KhmerFormer. This novel hybrid architecture intelligently combines the strengths of CNNs and Transformers. It uses multi-scale CNNs to extract robust local features from the glyph images and integrates an external attention mechanism into a Transformer encoder. This allows the model to capture not only fine-grained details but also crucial long-range dependencies and global features. This dual approach enables KhmerFormer to learn highly discriminative representations, setting a new state-of-the-art for classifying ancient Khmer glyphs.
A Low-Intervention Dual-Loop Iterative Process for Efficient Dataset Expansion and Classification in Palm Leaf Manuscript Analysis
N. Thuon, J. Du, ICDAR-IJDAR 2025.
Summary: The scarcity of labeled data is a major bottleneck in training accurate models. This work introduces a Low-Intervention Dual-Loop Iterative Process to overcome this. This semi-supervised method starts with a small, labeled seed set to train an initial model. In the first loop, the model makes predictions on a large pool of unlabeled data, and high-confidence predictions are automatically added to the training set. In the second, human-in-the-loop, an expert provides minimal intervention by correcting only the most uncertain predictions identified by the model. This efficient cycle rapidly and economically expands the dataset with high-quality labels, leading to significant improvements in classification performance with minimal manual effort.
How to Cite
Improving Isolated Glyph Classification Task for Palm Leaf Manuscripts
@inproceedings{thuon2022improving,
title={Improving isolated glyph classification task for palm leaf manuscripts},
author={Thuon, Nimol and Du, Jun and Zhang, Jianshu},
booktitle={International Conference on Frontiers in Handwriting Recognition},
pages={65--79},
year={2022},
organization={Springer}
}KhmerFormer: Multi-Scale CNNs-Transformer with External Attention for Ancient Khmer Isolated Glyph Classification
@inproceedings{thuon2024khmerformer,
title={Khmerformer: Multi-scale cnns-transformer with external attention for ancient khmer palm leaf isolated glyph classification},
author={Thuon, Nimol and Du, Jun},
booktitle={2024 Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC)},
pages={1--6},
year={2024},
organization={IEEE}
}A Low-Intervention Dual-Loop Iterative Process for Efficient Dataset Expansion and Classification in Palm Leaf Manuscript Analysis
@article{thuon2025low,
title={A Low-Intervention Dual-Loop Iterative Process for Efficient Dataset Expansion and Classification in Palm Leaf Manuscript Analysis},
author={Thuon, Nimol and Du, Jun and Theang, Panhapin and Thuon, Ratana},
journal={International Journal on Document Analysis and Recognition (IJDAR)},
pages={1--18},
year={2025},
publisher={Springer}
}