Autonomous Document Cleaning--A Generative Approach to Reconstruct Strongly Corrupted Scanned Texts

Area of Science:

• Computer Vision
• Document Image Analysis
• Machine Learning

Background:

• Scanned documents often suffer from severe degradation due to dirt, ink spills, and manual markings.
• Existing methods may require extensive training data or struggle with diverse corruption types.

Purpose of the Study:

• To develop an autonomous method for cleaning heavily corrupted scanned text documents.
• To recover clean text from a single page using only its inherent information.

Main Methods:

• Unsupervised learning of character representations from document patches using a probabilistic generative model.
• Efficient parameter inference via a truncated variational Expectation-Maximization (EM) algorithm.
• Pattern class and position identification using learned representations and a quality measure for discrimination.

Main Results:

• Demonstrated effective cleaning of documents with limited character types, relying on structural regularity.
• Showcased the approach's effectiveness, efficiency, and generality across different alphabets.
• Confirmed that a single page may not contain enough examples for a full alphabet but is sufficient for limited character sets.

Conclusions:

• The proposed method autonomously cleans corrupted documents by leveraging learned character structural regularities.
• This approach offers an efficient and general solution for document image restoration, particularly for pages with limited character diversity.
• The unsupervised learning framework provides a robust way to handle severe document degradation without manual supervision.