Document Scanner

A document scanner built with well-known computer vision methods in Python and OpenCV.

This project takes a photo of a document, detects the page boundary, corrects perspective distortion, and produces a high-contrast scanned result. It was developed as part of a computer vision course (EE146 @ UCR) and focuses on implementing the full pipeline from edge detection to perspective rectification and document enhancement.

Features

• Detects the document boundary from an image using contour analysis
• Scores candidate quadrilaterals based on area, convexity, side lengths, and corner angles
• Applies a four-point perspective transform to deskew the document
• Produces a binarized scan-like image using adaptive thresholding
• Supports optional CLAHE preprocessing for difficult lighting conditions
• Saves intermediate debugging artifacts such as grayscale, edges, closed edges, and detection overlays
• Includes both single-image testing and batch evaluation scripts

Dataset

This project uses the GNHK (GoodNotes Handwriting Kollection) dataset for evaluation and benchmarking of document boundary detection and perspective correction.

The GNHK dataset consists of unconstrained, camera-captured images of English handwritten documents collected from diverse real-world environments. It is designed to reflect realistic variations in lighting, orientation, background clutter, and handwriting styles, making it well-suited for evaluating document analysis and computer vision pipelines.

The dataset was introduced in:

Lee, A. W. C., Chung, J., & Lee, M. (2021). GNHK: A Dataset for English Handwriting in the Wild. ICDAR 2021.

• Official page: https://www.goodnotes.com/gnhk/
• GitHub repository: https://github.com/GoodNotes/GNHK-dataset

Usage

Uploading Images:

Place your image anywhere in the project (recommended: documents/raw/)

Single Image Scan:

python -m scripts.single_img_test --img <image-path> --outdir <output-dir>

Example:

python -m scripts.single_img_test --img documents/raw/base.jpg --outdir outputs/debug/single_img

Batch Processing:

python -m scripts.multi_img_test --input <input-dir> --outdir <output-dir>

Example:

python -m scripts.multi_img_test --input documents/raw  --outdir outputs/debug/eval

Future Work

• Deep Learning Document Detection (e.g. YOLO, U-Nets)
• Improved Corner Detection (e.g. learning-based keypoint detection, sub-pixel corner refinement)
• Mobile/Real-Time Deployment
• Intuitive User Interface

Acknowledgements

This project was developed as part of EE146 (Computer Vision) at the University of California, Riverside, under the instruction of Professor Bir Bhanu, with guidance from Ankith Jain Rakesh Kumar.