Angkorian-KSI-Small

Dataset Summary

Angkorian-KSI-Small, also referred to as KSI-B-Small, is a compact public sample derived from the Angkorian-KSI research project. It supports preliminary experiments in Khmer stone-inscription binarization, where a color photograph of a carved inscription is converted into a binary foreground–background representation.

Khmer stone inscriptions present a substantially different domain from scanned or handwritten documents. Their visual appearance is affected by shallow relief, erosion, cracks, biological growth, uneven illumination, cast shadows, surface deposits, complex stone texture, and overlapping carvings. These conditions make conventional document-binarization methods difficult to transfer directly.

This small release is intended for:

• testing data-loading and preprocessing pipelines;
• inspecting the image and annotation format;
• conducting small-scale or pilot binarization experiments;
• supporting reproducibility and responsible cultural-heritage research;
• demonstrating the domain gap between conventional document images and carved stone inscriptions.

Important: KSI-B-Small is a limited sample and is not representative of the size, diversity, or full evaluation scope of Angkorian-KSI. Results obtained on this release should be treated as preliminary and should not be presented as full-benchmark results.

Dataset Card

Item	Description
Dataset	Angkorian-KSI-Small (KSI-B-Small)
Domain	Khmer stone inscriptions
Primary task	Document image binarization / text-mask extraction
Language and script	Khmer
Training split	10 image–mask pairs
Test split	5 image–mask pairs
Total	15 image–mask pairs
Input	RGB photographs of stone inscriptions
Target	Binary text masks
Full project	Angkorian-KSI
Published paper	ICDAR 2026 paper

Relationship to Angkorian-KSI

The full Angkorian-KSI benchmark studies three complementary tasks:

1. KSI-LA — Layout Analysis: detection of inscription regions and text lines.
2. KSI-B — Binarization: extraction of carved-text masks from degraded stone surfaces.
3. KSI-C — Script-Period Classification: classification of inscriptions into historical periods.

KSI-B-Small includes only a small sample of the KSI-B task. It does not include the complete benchmark, the full annotation collection, or the complete official evaluation splits.

Dataset Structure

Each example should contain:

• an input photograph of a Khmer stone inscription;
• its corresponding binary ground-truth mask;
• a stable image identifier;
• a split designation (train or test).

A typical organization is:

Angkorian-KSI-Small/
├── train/
│   ├── original/
│   └── binarized/
└── test/
    ├── original/
    └── GT/

Input original and binarized should be matched using the same base filename. Users should verify the current repository structure when implementing custom data loaders.

Intended Use

KSI-B-Small is intended for responsible research and education related to:

• cultural-heritage documentation and preservation;
• Khmer and Southeast Asian historical-document analysis;
• low-resource document image analysis;
• image enhancement and binarization;
• domain adaptation and transfer learning;
• robustness analysis under erosion, texture, and illumination changes.

The release is not intended for commercial exploitation, reconstruction of sensitive site information, removal of cultural objects from their historical context, or claims about provenance, ownership, authenticity, translation, or historical interpretation without qualified expert review.

Cultural-Heritage and Access Notice

Angkorian-KSI concerns culturally significant Khmer heritage materials. Users should preserve cultural context, provenance, and scholarly attribution when using or presenting the data.

• KSI-B-Small is provided as a public sample for research, testing, education, and format inspection.
• The full Angkorian-KSI dataset and annotations are restricted to approved, non-commercial cultural-heritage research.
• Public availability of this sample does not grant access to the full benchmark.
• Redistribution, repackaging, or commercial use of the full restricted resources is not permitted without explicit authorization from the project team and relevant heritage stakeholders.
• Model outputs must not be treated as authoritative epigraphic readings, historical interpretations, or conservation assessments.

Researchers seeking access to the full benchmark should consult the official Angkorian-KSI repository for the current access procedure.

Experimental Protocol

The reported KSI-B-Small experiments are intentionally limited pilot studies.

• Classical methods are tuned using training data only.
• Zero-shot methods are evaluated without KSI-B-Small fine-tuning.
• DIBCO zero-shot models use weights trained on external document-binarization data.
• Fine-tuned methods use only the small KSI-B-Small training partition.
• Training configurations are limited to a maximum of 50 epochs.
• Runtime is reported in seconds per image under the project evaluation environment.
• Values are reported as mean ± standard deviation across test images.

Because the test partition contains only five images, the results have high uncertainty and should be interpreted as diagnostic evidence rather than definitive model rankings.

Preliminary Baseline Results

Higher values are better for F-measure, pseudo-F-measure, and PSNR. Lower values are better for DRD and seconds/image.

Method	Family	Track	F ↑	pseudo-F ↑	PSNR ↑	DRD ↓	seconds/image ↓
Otsu	classical/global	train-tuned	0.155 ± 0.100	0.146 ± 0.113	2.867 ± 1.211	102.426 ± 34.817	0.0077 ± 0.0029
Sauvola	classical/local	train-tuned	0.141 ± 0.093	0.134 ± 0.121	6.163 ± 1.838	47.394 ± 25.265	0.0190 ± 0.0066
Wolf–Jolion	classical/local	train-tuned	0.140 ± 0.069	0.128 ± 0.094	6.367 ± 1.463	43.138 ± 17.313	0.0128 ± 0.0037
SAE/DAE	autoencoder	zero-shot	0.219 ± 0.152	0.205 ± 0.181	5.202 ± 2.326	62.703 ± 36.833	0.6111 ± 0.1908
DP-LinkNet	CNN	zero-shot	0.088 ± 0.149	0.097 ± 0.176	8.553 ± 0.941	22.851 ± 6.053	0.4786 ± 0.1635
U-Net	CNN	fine-tuned	0.432 ± 0.140	0.452 ± 0.146	8.070 ± 0.973	26.704 ± 7.621	0.7294 ± 0.3370
SAE/DAE	autoencoder	fine-tuned	0.162 ± 0.138	0.164 ± 0.159	6.582 ± 1.497	42.426 ± 21.570	0.6135 ± 0.1927
DocDiff	diffusion	DIBCO zero-shot	0.143 ± 0.117	0.139 ± 0.137	5.129 ± 1.606	59.892 ± 25.102	5.6917 ± 1.5151
DocEnTr	transformer	zero-shot	0.130 ± 0.120	0.129 ± 0.142	8.378 ± 1.505	24.756 ± 10.358	4.3325 ± 1.3583
U-Net	CNN	DIBCO zero-shot	0.134 ± 0.139	0.131 ± 0.157	8.538 ± 1.226	23.329 ± 8.245	0.7307 ± 0.2538
DocDiff	diffusion	fine-tuned	0.138 ± 0.025	0.136 ± 0.024	6.562 ± 0.476	38.300 ± 6.250	5.4655 ± 1.3868
DE-GAN	GAN	zero-shot	0.255 ± 0.179	0.247 ± 0.194	3.807 ± 2.385	89.525 ± 44.531	2.0661 ± 0.6800
DE-GAN	GAN generator	fine-tuned	0.487 ± 0.148	0.497 ± 0.153	7.769 ± 1.037	29.575 ± 9.899	1.9814 ± 0.6272
PALM-GAN	CNN–Transformer hybrid	raw experimental	0.476 ± 0.173	0.486 ± 0.168	7.779 ± 1.005	29.409 ± 9.681	1.9589 ± 0.6166
PALM-GAN	CNN–Transformer hybrid	fine-tuned	0.508 ± 0.165	0.519 ± 0.169	8.479 ± 0.904	26.762 ± 8.712	1.9495 ± 0.5920

PALM-GAN obtains the highest preliminary F-measure and pseudo-F-measure, while DP-LinkNet obtains the strongest PSNR and DRD values. These metric differences indicate that no single method dominates every aspect of binarization quality. Given the very small test set, these observations should not be generalized to the full benchmark.

Evaluation Metrics

• F-measure (F): pixel-level balance between foreground precision and recall.
• pseudo-F-measure (pseudo-F): skeleton-aware evaluation that emphasizes preservation of text structure.
• Peak Signal-to-Noise Ratio (PSNR): pixel-level similarity between the predicted mask and ground truth.
• Distance Reciprocal Distortion (DRD): perceptually weighted binary-image distortion; lower is better.
• seconds/image: average inference time per image in the reported environment.

Limitations

• The release contains only 15 image–mask pairs.
• The five-image test set cannot provide stable estimates of general performance.
• Mean and standard deviation values may be strongly influenced by individual samples.
• The sample does not cover the full diversity of temples, surfaces, degradation patterns, script periods, capture conditions, or annotation cases represented by the complete research project.
• Models trained on this release may overfit and should not be assumed to generalize to unseen sites or inscription styles.
• Binary masks simplify complex and sometimes ambiguous carved surfaces; expert interpretation may still be required.
• Results are not directly comparable with experiments using different splits, preprocessing, resolution, hardware, checkpoints, or thresholding procedures.

Responsible Use

Users should:

• cite the Angkorian-KSI paper and dataset repository;
• describe the release explicitly as KSI-B-Small rather than the full benchmark;
• report the exact split, preprocessing, model checkpoint, and training configuration;
• avoid publishing sensitive location or provenance information not already made public by authorized heritage institutions;
• consult qualified Khmer epigraphers, archaeologists, historians, or conservators for cultural and historical claims;
• clearly communicate uncertainty and known failure cases;
• avoid using generated masks as a substitute for professional conservation decisions.

Method References

• DP-LinkNet: DP-LinkNet: A convolutional network for historical document image binarization
• DE-GAN: DE-GAN: A Conditional Generative Adversarial Network for Document Enhancement
• DocEnTr: DocEnTr: An End-to-End Document Image Enhancement Transformer
• DocDiff: DocDiff: Document Enhancement via Residual Diffusion Models
• PALM-GAN: Generate, transform, and clean: the role of GANs and transformers in palm leaf manuscript generation and enhancement

Citation

If you use KSI-B-Small, please cite the Angkorian-KSI paper:

The paper was first published online on 24 August 2026 as part of ICDAR 2026; Springer’s official bibliographic record uses the publication year 2027, which is retained below.

@inproceedings{thuon2027angkorian,
  title     = {Angkorian-KSI: A Multi-task Benchmark for Khmer Stone Inscription Analysis},
  author    = {Thuon, Nimol and Du, Jun and Thuon, Ranysakol and Theang, Panhapin},
  booktitle = {Document Analysis and Recognition -- ICDAR 2026},
  series    = {Lecture Notes in Computer Science},
  volume    = {16974},
  pages     = {387--404},
  year      = {2027},
  publisher = {Springer Nature Switzerland},
  doi       = {10.1007/978-3-032-36039-7_23},
  url       = {https://doi.org/10.1007/978-3-032-36039-7_23}
}

Please also identify the data release used in your work:

Angkorian-KSI-Small (KSI-B-Small), accessed from
https://huggingface.co/datasets/Backkh/Angkorian-KSI-Small

Links

• Dataset: https://huggingface.co/datasets/Backkh/Angkorian-KSI-Small
• Code and project documentation: https://github.com/back-kh/Angkorian-KSI
• Project website: https://angkorianai.github.io/
• Published paper: https://link.springer.com/chapter/10.1007/978-3-032-36039-7_23

Contact

Project lead, Nimol Thuon, at nimol.thuon@gmail.com, or use the contact information provided in the official Angkorian-KSI repository.

---

license: cc-by-nc-nd-4.0