Inscription intelligence
Recovering structure from carved text.
Detect text regions and lines, extract binary text masks, and recognize historical periods across severely weathered Khmer stone inscriptions.
KSI–LAKSI–BKSI–C
Menu
Angkorian–AI ProjectJan 2026 — Present
Computational research for Khmer stone heritage.
Angkorian–AI advances computer vision for Khmer stone heritage— building datasets, benchmarks, and preservation-centered methods for inscription analysis and visual condition assessment.
FieldComputer vision & document analysis
DomainKhmer stone heritage
Research outputs2 conference papers · 2026
Primary regionAngkor, Cambodia
Research program / 01
One preservation mission. Two complementary computational views.
AI that begins with the material evidence.
Khmer stone surfaces are neither clean pages nor standard objects. Relief-induced shadows, erosion, biological growth, texture, and historical variation shape every task we study.
Condition intelligence
Making visible degradation measurable.
Support expert assessment of stone damage and severity with classification, bounding-box detection, segmentation, and interpretable visual evidence.
DetectionSegmentationAssessment
Selected publications / 02
Conference proceedings · 2026
Publications
Two complementary studies establish the project’s research agenda across historical document analysis and visual condition assessment.
Conference paper · First online 24 Aug 2026
Angkorian-KSI: A Multi-task Benchmark for Khmer Stone Inscription Analysis
DOI: 10.1007/978-3-032-36039-7_23
The first benchmark dedicated to automated digital analysis of Khmer stone inscriptions, spanning structural detection, text-mask extraction, and historical period recognition. Baselines expose a pronounced domain gap created by erosion, relief shading, complex stone texture, and linguistic variation.
Conference workshop paper · Aug 2026
Angkorian-HeritageObj: AI-Assisted Visual Condition Assessment of Khmer Stone Heritage Objects
ICPR–PRESTIGE Workshop · August 2026
A preservation-oriented study of AI-assisted visual condition assessment for Khmer stone objects, evaluating modern vision architectures against damage severity, surface degradation, and demanding field-capture conditions.
Angkorian-KSI / Benchmark
A compact dataset built for a difficult domain.
Curated from in situ captures across multiple sites in a UNESCO World Heritage archaeological region, Angkorian-KSI turns field imagery into a coordinated three-task evaluation framework.
View paper details230full inscription images
760annotated text regions
2,733annotated text lines
3,493binarization masks
534script-period labels
KSI-B-Small public release
| Resource | Task | Split | Status | Access |
|---|---|---|---|---|
| KSI-B-Small | Khmer stone-inscription binarization | 10 training / 5 test images | Public | Hugging Face ↗ · GitHub ↗ |
Research-use notice: the full Angkorian-KSI benchmark remains restricted to approved, non-commercial cultural-heritage research. KSI-B-Small is provided as a public sample for testing and format inspection.
How we work / 03
A preservation-centered research pipeline
From field capture to evidence.
Each stage is designed to keep provenance, uncertainty, and expert review visible.
- 01
Document
Capture objects, inscriptions, surfaces, and context under real field conditions.
- 02
Curate
Clean and structure images with site, period, object, and provenance metadata.
- 03
Annotate
Build detailed polygons, masks, bounding boxes, text lines, and expert-defined labels.
- 04
Benchmark
Evaluate modern vision architectures against Khmer stone’s distinctive domain shift.
- 05
Return
Translate outputs into inspectable evidence for researchers and heritage specialists.
Inside the research / 04
The benchmark makes failure visible.
Severe texture, erosion, lighting, and overlapping carvings test the limits of systems trained on cleaner visual domains.
Project leadership / 05
Applied AI · Historical analysis · Low-resource languages
Nimol Thuon, PhD
Project Director · Research Scientist
Dr. Nimol Thuon is a research scientist specializing in applied AI for historical analysis and low-resource languages. For more than a decade, his research has examined multi-script historical manuscripts across South and Southeast Asia, covering 8–10 writing systems and a broad range of low-resource languages. He now leads Angkorian–AI, advancing computational analysis of Khmer stone inscriptions and heritage objects for preservation, documentation, and long-term cultural research.
10+ years of research8–10 writing systemsSouth & Southeast AsiaKhmer heritage preservation
Partners & supporters
Institutional partnership and research support
Collaboration
Let’s build a responsible digital future for Khmer heritage.
We welcome conversations with conservators, archaeologists, epigraphers, museums, Cambodian universities, and computer vision researchers working on preservation-centered methods.
Contact the project