CHULA

01 · Overview

Research Motivation

Segmentation of legal cadastral documents remains extremely challenging under severe real-world degradation.

Real-world land title deeds are noisy, degraded, and legally sensitive.

Existing systems fail under handwriting overlays, faded scans, official stamps, blur, and layout corruption.

Even minor recognition errors can produce critical legal inconsistencies in national-scale document digitization pipelines.

CHULA introduces a custom heuristic uncertainty-guided loss for robust land title deed segmentation under real operational constraints.

Overview Figure — CHULA segmentation pipeline

02 · Methodology

Core Architecture

Robust segmentation through heuristic uncertainty-guided loss formulation.

01 ──

🎯

Uncertainty Estimation

Detect unreliable visual regions under severe degradation, handwriting contamination, and scanning artifacts to guide adaptive loss weighting.

02 ──

⚡

Heuristic Loss Guidance

Custom loss function that dynamically suppresses noisy regions including blur, occlusion, official stamps, and structural corruption during training.

03 ──

🔐

Structured Segmentation

Preserve cadastral boundary consistency for reliable downstream legal verification and national-scale document processing.

03 · Benchmarks

Benchmark Results

CHULA consistently outperforms existing baselines across challenging cadastral document benchmarks.

Method	OCR Accuracy ↑	NED ↓	Cadastral Match ↑
CRNN	61.3	5.42	48.1
TrOCR	71.4	4.82	57.2
PARSeq	74.1	3.94	63.5
CHULA (Ours) ★ SOTA	92.6	1.64	91.7

04 · Demonstration

System Demo

Real-world segmentation inference under severe cadastral degradation.

05 · Impact

Real-World Impact

Designed for national-scale public infrastructure deployment.

🏛

National Land Registry

Enable scalable cadastral digitization for government administrative systems across Thailand and beyond.

⚖

Legal Verification

Improve reliability and consistency in official legal document processing pipelines.

📑

Administrative Automation

Reduce manual workload in large-scale document archival and processing pipelines.

🌏

Public AI Infrastructure

Support AI-driven digital transformation for governmental services at national scale.

06 · Visual Results

Qualification Results

Real Thai land title deeds — processed at national scale.
Every bounding box is a field extracted by CHULA under extreme real-world degradation.

🌀

Degraded Scans

Decades-old documents with fading ink, yellowing paper, and scan noise

✍️

Handwriting Overlay

Manual annotations and stamps blending with printed cadastral text

🗺️

Complex Layouts

Multi-field structured forms with irregular boundary coordinates

⚖️

Zero Tolerance

Legal documents — a single wrong field can invalidate land ownership records

CHULA detection result on Thai land title deed sample 01

SAMPLE_01 ↑ High Confidence Fields

01 ──

Cadastral Boundary & PIN Detection

A degraded Chiang Rai land deed with overlapping handwritten marks and an official circular stamp. CHULA successfully localizes PIN fields, side-land measurements, north arrow, export date and year — all under severe visual noise. Note the high confidence scores even on occluded fields.

PIN_OCR SIDE_LANDS_OCR NORTH_OBB EXPORT_DATE EXPORT_YEAR

CHULA detection result on Thai land title deed sample 02

SAMPLE_02 24+ Fields Detected

02 ──

Complex Parcel Layout — Multiple Neighbor Fields

A Chiang Rai parcel with an irregular polygon boundary and dense text distribution around the perimeter. CHULA detects PIN nodes along each boundary segment, land-scale metadata, and neighbor parcel labels (SIDE_LANDS_OCR) with robust localization despite layout complexity.

PIN_OCR ×24 LANDS_SCALE OBJECT_OVERLAY NORTH_OBB EXPORT_YEAR 0.81

CHULA detection result on Thai land title deed sample 03

SAMPLE_03 40+ Fields Detected

03 ──

High-Density Field Detection — Urban Parcel

The most challenging case: a dense urban parcel map with dozens of PIN coordinate fields, multiple SIDE_LANDS neighbor annotations, and a partially visible north arrow. CHULA achieves high-confidence detections across 40+ fields simultaneously, demonstrating scalability to the most complex cadastral documents in the national archive.

40+ Fields OBJECT_LANDS_OCR SIDE_LANDS_OCR NORTH_OBB 0.91 EXPORT 0.88

07 · Funding

Acknowledgement

This research was supported by a prestigious postdoctoral fellowship enabling frontier AI for national infrastructure.

Research Grant · PostDoc Fellowship

The Second Century Fund (C2F)
PostDoctoral Fellowship

Chulalongkorn University, Thailand

This work was supported by the C2F (The Second Century Fund) PostDoctoral Fellowship Program at Chulalongkorn University. C2F accelerates frontier AI research bridging academic excellence with national-scale real-world deployment — enabling CHULA to move from laboratory research to public digital infrastructure serving all of Thailand.

30M+ Land deeds targeted

IEEE Access 2025

92.6% OCR accuracy

🇹🇭 National impact

CHULA demonstrates that robust segmentation for legally sensitive documents requires more than stronger recognition models — it requires uncertainty-aware loss formulation grounded in operational constraints and real-world document characteristics.

08 · Citation

BibTeX

Cite our work if you find it useful for your research.

bibtex · panboonyuen2025chula

@article{panboonyuen2025chula,
  title   = {CHULA: Custom Heuristic Uncertainty-Guided Loss for Accurate Land Title Deed Segmentation},
  author  = {Panboonyuen, Teerapong},
  journal = {IEEE Access},
  year    = {2025},
  publisher = {IEEE}
}