Document Scanner from Scratch

A pure Python document scanner using only NumPy, Numba (to speed up execution) developed from scratch for academic purposes.

CI/CD	Python	License

Overview

This project is a document scanner built from scratch in Python.

It was developed as part of my TIPE (Personal Initiative Project) in engineering school. The goal was to build a project entirely by ourself, without relying on high-level libraries. This approach allowed me to deeply understand the underlying algorithms—from edge detection to perspective correction—and implement them step by step. It's a testament to learning by doing and the value of reinventing the wheel for educational purposes.

Goal: Archieve an automatic scan of a document. Document will be straightened and binarized in dark and white.

Features

Feature	Description
Corner Detection	Canny Edge detection
Perspective Correction	Homographic transformation to straighten the document
Binarization	Convert to black and white using local thresholding
Export	Save as PNG (matplotlib)

Installation

Prerequisites

• Python 3.8+
• Jupyter Notebook
• Matplotlib
• NumPy
• Pil (import image only)
• Numba (optional)

Installation Steps

git clone https://github.com/TooLoss/PythonDocumentScanFromScratch.git
cd PythonDocumentScanFromScratch
pip install -r requirements.txt

Usage

Every steps are details in the main notebook.

Methodology

1. Corner Detection

• Sobel filters to detect edges.
• Hadamar product and Connected Component Labeling to find corners.

2. Perspective Correction

• Calculate homographic transformation using the detected corners.
• Apply transformation and fill unaffected values with a nearest neighbors algorithm.

3. Binarization

• Local tresholding implementation : Sauvola and Niblack.

Jupyter Notebooks

This project includes Jupyter Notebooks for exploration and results visualization:

Notebook	Description
main.ipynb	Automatic scan process.
/notebook/OptimisationSortie.ipynb	How to optimize the size of the projected document.
	Other notebooks have to be cleaned up.

Results and limits

The process works best if it's a black and white text document. With a high contrast between the page and the background.

Metric	Value
Detection Accuracy (with default parameters)	Testing ...
Execution Time : Projective transofrm	12s per image
Execution Time : Tresholding	1m per image

Contributing

Contributions are welcome! Open an Issue or a Pull Request to suggest improvements.

License

This project is licensed under the MIT License. See LICENSE for details.

Academic References

FILIPPO BERGAMASCO : Computer Vision : Projective geometry and 2D transformations / Projective Transform

DJEMEL ZIOU : La détection de contours dans des images à niveaux de gris : mise en œuvre et sélection de détecteurs / Edge detection

L. JAGANNATHAN ; C. V. JAWAHAR : Perspective Correction Methods for Camera- Based Document Analysis / Homography Matrix

Contact

Bilèle EL HADDADI

LinkedIn | GitHub

---

© 2025 – Project developed as part of the TIPE.