AI-Based Ukrainian Handwritten Document Recognition System

End-to-end OCR + HTR pipeline for the Kaggle Handwritten to Data 2026 hackathon using the RUKOPYS handwritten document dataset.

This repository contains a complete project scaffold:

• Document preprocessing
• Layout/text-region detection
• Region classification
• OCR/HTR recognition
• Ukrainian language correction hooks
• Kaggle submission generation
• Evaluation utilities
• Streamlit demo app
• Hackathon documentation and architecture notes

Project Structure

.
├── app/
│   └── streamlit_app.py
├── configs/
│   └── default.yaml
├── data/
│   ├── raw/
│   ├── processed/
│   └── submissions/
├── docs/
│   ├── architecture.md
│   └── hackathon_report.md
├── scripts/
│   ├── run_inference.py
│   └── train_baseline.py
├── src/
│   └── rukopys_ocr/
│       ├── config.py
│       ├── correction.py
│       ├── detection.py
│       ├── metrics.py
│       ├── pipeline.py
│       ├── preprocessing.py
│       ├── recognition.py
│       ├── schemas.py
│       └── submission.py
└── tests/
    ├── test_metrics.py
    └── test_pipeline_smoke.py

Quick Start

Create an environment and install dependencies:

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt

Run the demo app:

streamlit run app/streamlit_app.py

Run batch inference:

python scripts/run_inference.py --input data/raw/test_images --output data/submissions/submission.csv

Run tests:

pytest

Dataset Layout

Place Kaggle or HuggingFace-exported data in data/raw/:

data/raw/
├── train_images/
├── test_images/
├── train.jsonl
├── silver.jsonl
└── sample_submission.csv

The pipeline accepts common image formats: .png, .jpg, .jpeg, .tif, .tiff, .bmp.

Baseline Approach

The default implementation is intentionally modular. It includes a lightweight classical fallback so the project runs without downloading large models, while leaving clear extension points for production models:

• Detector: contour-based fallback, replaceable with YOLO/Detectron2
• Recognizer: optional Tesseract/PaddleOCR-style adapter, replaceable with TrOCR/Qwen/Gemma VLM
• Corrector: dictionary and normalization hooks for Ukrainian post-processing
• Submission: Kaggle-ready CSV generation

Competition Goal

Optimize the combined score:

0.15 * DetectionF1
+ 0.05 * ClassificationAccuracy
+ 0.30 * (1 - CER)
+ 0.50 * (1 - PageCER)

Prepared By

Shivam Singh
Department of Computer Science Engineering
AI Research and Development
2026