SEDS536 Exam Paper Reader & OCR

This project segments a raw exam photo to isolate the paper, runs OCR models on the text regions, and evaluates predictions against labeled answers.

Pipeline (Short)

1. Segment the page (shadow correction + threshold/edge masks + contour crop).
2. Detect text regions with PaddleOCR.
3. Crop each region and run PaddleOCR + EasyOCR + TrOCR (base/large) for recognition.
4. Save raw + cleaned predictions and compute CER/accuracy per model.

Pipeline Diagram

Methodology

1. Page segmentation: Normalize illumination, build mask/edge cues, and extract the page contour to crop a clean paper view.
2. Text region detection: Use PaddleOCR detection to locate line-level text polygons on the cropped page.
3. Region normalization: Warp each polygon to a rectangular crop so OCR models see upright text.
4. Multi-model OCR: Run PaddleOCR, EasyOCR, and TrOCR (base/large) on each crop for complementary predictions.
5. Answer structuring: Detect question numbers (1–20) per line, anchor bbox method per question.
6. Evaluation: Compare model answers with ground-truth labels to compute per-model accuracy and CER.

Project Structure

exam-paper-reader-ocr/
|- README.md
|- requirements.txt
|- demo.py             # Demo runner (writes artifacts to demo/)
|- demo/               # Demo outputs (step images + metrics)
|- data/                # Input images
|- predictions/         # Raw/cleaned predictions + metrics
|- segment-data/        # Saved segmentation steps
|- docs/
   |- ground-truth-labels/
      |- answers.json   # Labeled answers for evaluation
|- src/
   |- main.py           # End-to-end pipeline
   |- segmentation.py   # Page segmentation + step images
   |- models.py         # PaddleOCR, EasyOCR, TrOCR

Run

1. Create venv: python -m venv venv
2. Activate: source venv/bin/activate
3. Install deps: pip install -r requirements.txt
4. Run: python src/main.py

Default image path is set in src/main.py (IMAGE_PATH). Outputs are written to predictions/ and step images to segment-data/.

Batch Run

Run the pipeline for every image in data/:

python - <<'PY'
from pathlib import Path
import sys

sys.path.insert(0, "src")
import main

data_dir = Path("data")
for img in sorted(data_dir.glob("*.png")):
    print(f"Processing {img.name}")
    main.main(str(img))
PY

Demo

Run the end-to-end demo and collect artifacts in demo/:

python demo.py

Optional: pass a specific image path:

python demo.py data/exam68.png

The demo folder will include segmentation step images and the run metrics JSON for quick presentation.