Document ink bleed-through removal with two hidden Markov random fields and a single observation field

Area of Science:

• Computer Vision
• Image Processing
• Digital Humanities

Background:

• Document bleed-through is a common issue in historical documents, hindering readability and automated analysis.
• Existing methods often struggle with accurately separating superimposed text from recto and verso pages.

Purpose of the Study:

• To develop an effective blind method for removing document bleed-through.
• To improve the accuracy of text recognition on degraded historical documents.

Main Methods:

• A novel approach utilizing separate Markov Random Field (MRF) regularization for recto and verso sides.
• Bayesian Maximum a Posteriori (MAP) estimation for segmentation.
• Optimization via minimum cut/maximum flow algorithm for binary labeling problems.

Main Results:

• The proposed method demonstrates improved pixel estimation by considering recto and verso page interactions.
• Evaluated on 18th-century scanned documents, it shows enhanced character recognition accuracy compared to existing restoration techniques.

Conclusions:

• The separate MRF regularization approach effectively addresses document bleed-through.
• This technique offers significant improvements for historical document analysis and digital preservation.