# End-to-End Unsupervised Document Image Blind Denoising

**Authors:** Mehrdad J Gangeh, Marcin Plata, Hamid Motahari, Nigel P Duffy  
**arXiv:** [2105.09437](https://arxiv.org/abs/2105.09437)  
**Date:** 2021-05-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2105.09437v2/

## Key points

- Authors address the task of removing noise from scanned pages before OCR
- Limitations of existing supervised denoising methods due to lack of noisy/clean page pairs
- Lack of a single model for effectively removing various types of noise from documents
- Proposal of an end-to-end unsupervised deep learning model for document image denoising
- Model designed to remove salt & pepper noise, blurred/faded text, and watermarks at different intensities
- Extensive testing shows significant enhancement in image quality and OCR performance
- Research introduces a comprehensive solution without requiring labeled data for training

## Summary

In their paper titled "End-to-End Unsupervised Document Image Blind Denoising," authors Mehrdad J Gangeh, Marcin Plata, Hamid Motahari, and Nigel P Duffy address the crucial task of removing noise from scanned pages before submitting them to optical character recognition (OCR) systems. They highlight the limitations of existing supervised image denoising methods that require pairs of noisy/clean pages. These pairs are often not readily available in real-world settings. Furthermore, they emphasize the lack of a single model capable of effectively removing various types of noise from documents. To tackle these challenges, the authors propose a groundbreaking unified end-to-end unsupervised deep learning model. This innovative model is designed to remove multiple types of noise from scanned documents including salt & pepper noise, blurred or faded text, and watermarks at different levels of intensity. Through extensive testing on various datasets, the authors demonstrate that their proposed model significantly enhances the quality of scanned images and improves OCR performance on scanned pages. The research conducted by Gangeh et al. represents a significant advancement in the field of document image denoising by introducing a comprehensive solution that addresses multiple types of noise without requiring labeled data for training. Their findings have implications for improving document processing workflows and enhancing the accuracy and efficiency of OCR systems in handling noisy scanned documents.