status: in progress#khmer ocr#benchmark dataset#vision transformer (trocr)#crnn-ctc

khmer handwritten ocr on classroom whiteboards

Robust text recognition under realistic classroom conditions: glare, uneven lighting, and stacked Khmer script characters.

kirirom institute of technology·phnom penh, cambodia·chan ekmongkol
code ↗
executive overview & impact

Solves a long-standing barrier in Cambodian education: automatically converting classroom whiteboard notes into structured digital text without manual transcription or expensive cloud APIs.

key impact for education & digitisation:
  • ·Enables instant classroom digitization and homework generation for schools across Cambodia.
  • ·Handles extreme real-world challenges: mobile camera angles, marker fading, and high whiteboard glare.
  • ·Preserves Khmer handwriting nuance without losing vertical subscript vowels and consonant stacks.
research objective
To develop and evaluate an OCR framework for Khmer handwritten whiteboard text by creating a benchmark dataset and improving recognition under realistic classroom conditions.
scope & evaluation target
target domain
classroom whiteboards
architectures
trocr & crnn-ctc
evaluation
cer, wer & latency
dataset focus
benchmark & augmentation
research questions (rq)
RQ1

How can a benchmark dataset of Khmer handwritten whiteboard text be created for OCR development and evaluation?

RQ2

How can an OCR model be adapted to accurately recognize Khmer handwriting under classroom lighting and angle variations?

RQ3

How do preprocessing, model fine-tuning, and data augmentation affect OCR recognition performance?

RQ4

How much does this framework improve recognition accuracy and speed compared to existing OCR approaches?

core contributions
01.

benchmark dataset

  • Collect and annotate real-world Khmer handwritten whiteboard images from classrooms.
  • Structure the dataset for standardized training, validation, and testing.
02.

adapted ocr framework

  • Multi-stage image preprocessing: adaptive glare reduction and shadow removal.
  • Fine-tuned CRNN-CTC baseline and Vision Transformer (TrOCR) architectures.
  • Specialized data augmentation tailored to Khmer cursive writing variations.
03.

experimental evaluation

  • Ablation studies testing preprocessing, model tuning, and data augmentation.
  • Comparative benchmarks against standard OCR baselines.
  • Detailed metrics: Character Error Rate (CER), Word Error Rate (WER), and latency.
methodology

We build a dual-branch OCR pipeline designed for the unique challenges of Khmer script. The pipeline combines adaptive image filtering (to remove whiteboard reflections and lighting gradients) with transfer learning on recurrent (CRNN-CTC) and transformer (TrOCR) architectures trained on augmented real and synthetic handwriting data.

experimental roadmap
  • Dataset construction: Capturing and segmenting classroom whiteboard text across diverse markers, lighting angles, and handwriting styles.
  • Model training: Fine-tuning CRNN-CTC and TrOCR checkpoints with specialized Khmer character tokenization.
  • Ablation benchmarking: Systematically evaluating the impact of glare removal, elastic augmentation, and patch embeddings.
  • Edge optimization: Exporting quantized models to ONNX for fast inference on low-cost classroom devices.
technologies & framework stack
pytorchtrocrcrnn-ctcopencvhugging facetransformersfastapionnx
research questions or collaboration?
open for empirical research partnerships, dataset sharing, and technical discussion.