Computer Vision Document Scanner

An OpenCV-based document scanning pipeline that transforms photos of A4 pages, receipts, and notes into clean, top-down, high-contrast scans suitable for printing and OCR preprocessing.

Overview

This project detects documents in real-world photographs, corrects perspective distortion, reduces shadows, improves contrast, binarizes the result, and evaluates the quality of each generated scan.

The pipeline is designed to handle:

• cluttered or non-uniform backgrounds;
• documents photographed from imperfect angles;
• shadows and uneven lighting;
• low-contrast document boundaries;
• blurry or highly skewed inputs;
• single-image and folder-based batch processing.

Low-quality inputs can be automatically rejected based on sharpness, skew, and binarization quality.

Key Features

• Robust multi-pass document detection
• Perspective correction using a four-point homography
• De-shadowing and local contrast enhancement with CLAHE
• Otsu and adaptive thresholding
• Automatic selection of the preferred binary scan
• Sharpness estimation using the variance of the Laplacian
• Skew estimation using Hough lines
• Automatic rejection of low-quality inputs
• Batch processing for folders of images
• CSV reports containing detection, quality, and experiment results
• Visualization of intermediate processing stages

Processing Pipeline

flowchart TB
    A[Input image or folder] --> B[Resize and grayscale]
    B --> C[Document detection]
    C --> D[Four-corner contour]
    D --> E[Perspective rectification]
    E --> F[De-shadowing and CLAHE]
    F --> G[Otsu and adaptive thresholding]
    G --> H[Preferred binary scan]
    E --> I[Sharpness and skew metrics]
    H --> J[Binarization quality]
    I --> K{Quality gate}
    J --> K
    K -->|Accepted| L[Save scan and CSV results]
    K -->|Rejected| M[Save rejection reason]

Loading

How It Works

1. Pre-processing

The input image is resized for faster processing and converted to grayscale.

2. Document Detection

The system uses multiple detection strategies to locate a valid four-point document boundary:

1. bilateral filtering, Canny edge detection, and morphological closing;
2. adaptive thresholding and morphology;
3. Otsu-based blob detection;
4. Hough lines and corner detection;
5. minimum-area rectangle fallbacks.

Candidate contours are evaluated using properties such as area, convexity, aspect ratio, and rectangular fill ratio.

3. Perspective Rectification

The four detected corners are ordered and used to compute a perspective transformation.

OpenCV's getPerspectiveTransform and warpPerspective functions produce a top-down view of the document.

4. Image Enhancement

The rectified image is improved using:

• de-shadowing;
• CLAHE local contrast enhancement;
• grayscale normalization.

5. Binarization

Two methods are applied and compared:

• Otsu thresholding — effective for relatively uniform lighting;
• Adaptive thresholding — more robust to shadows and local lighting variation.

The pipeline selects the preferred result based on the image characteristics and binarization quality.

6. Quality Assessment

Each scan is evaluated using:

• Sharpness — variance of the Laplacian;
• Skew angle — estimated from Hough lines;
• Binarization quality — proportion and distribution of black pixels;
• Decision — Accepted or Rejected;
• Rejection reason — for example, low sharpness, excessive skew, poor binarization, or missing document contour.

Technologies

Technology	Purpose
Python	Main programming language
OpenCV	Image processing, contours, Hough transforms, homography, and thresholding
NumPy	Geometric calculations and image statistics
Matplotlib	Visualization of intermediate pipeline stages and experiment charts
CSV / OS	Batch processing and result reporting

Project Structure

Computer-Vision-Document-Scanner/
├── dataset/                       # Input document photographs
├── docs/
│   └── images/
│       └── example_result.png    # Example shown in this README
├── outputs_week9/                 # Generated scans, logs, and CSV reports
├── main.py                        # Main scanning pipeline
├── generate_charts.py             # Experiment chart generation
├── binarization_pie_chart.png
├── success_rate_by_lighting.png
├── requirements.txt
├── Documentatie.pdf               # Final project report
├── prezentare.pdf                 # Final presentation
└── README.md

Installation

1. Clone the repository

git clone https://github.com/AndreeaStati/Computer-Vision-Document-Scanner.git
cd Computer-Vision-Document-Scanner

2. Create a virtual environment

Linux or macOS

python3 -m venv .venv
source .venv/bin/activate

Windows

python -m venv .venv
.venv\Scripts\activate

3. Install the dependencies

pip install -r requirements.txt

Usage

1. Add .jpg, .jpeg, or .png document photos to the dataset/ directory.
2. Run the application:

python main.py

During visualization:

• press Enter to continue to the next image;
• press Q or Escape to stop processing.

Generated Outputs

For each detected document, the pipeline can generate files such as:

<image>_warped.jpg
<image>_deshadow.jpg
<image>_otsu.jpg
<image>_adaptive.jpg
<image>_preferred.jpg

The batch-processing workflow also produces CSV reports, including:

summary_week9.csv
experiments_week9.csv

The reports contain information such as:

• input filename;
• detection method;
• preferred binarization method;
• sharpness score;
• skew angle;
• lighting and background category;
• acceptance decision;
• rejection reason.

Dataset and Evaluation

The project was developed and evaluated on a dataset of 50 real-world document photographs captured under different:

• camera angles;
• lighting conditions;
• backgrounds;
• shadow levels;
• document types.

The evaluated methods include:

• simple Canny detection versus multi-pass detection;
• Otsu versus adaptive thresholding;
• enhancement with and without de-shadowing and CLAHE;
• output generation with and without automatic quality rejection.

Reported Results

On the reported demonstration subset:

Metric	Result
Document detection	9 / 9 images
Accepted scans	7 / 9 images
Acceptance rate	77.8%

Most rejected inputs were affected by motion blur, low sharpness, excessive skew, or weak binarization.

Example Result

The example below illustrates the main stages of the pipeline:

1. original photograph;
2. detected edges or document mask;
3. detected document contour;
4. perspective-corrected image;
5. enhanced image;
6. final binarized scan;
7. quality assessment and acceptance decision.

Possible Improvements

• OCR integration for automatic text extraction
• Automatic document orientation correction
• More advanced illumination normalization
• Deep-learning-based document boundary detection
• Mobile or web interface
• Evaluation on a larger and more diverse dataset

Academic project developed for the Artificial Vision Systems course.