Phase-based binarization of ancient document images: model and applications.

Area of Science:

• Computer Vision
• Image Processing
• Digital Humanities

Background:

• Binarization of ancient documents is crucial for digital preservation and analysis.
• Existing methods struggle with diverse degradation patterns common in historical manuscripts.
• Accurate ground truth generation is a bottleneck in developing and evaluating binarization techniques.

Purpose of the Study:

• To propose a novel phase-based binarization model for ancient document images.
• To introduce a versatile postprocessing technique to enhance existing binarization methods.
• To develop an efficient tool for generating ground truth data for historical documents.

Main Methods:

• Utilized three phase-derived feature maps: maximum moment of phase congruency covariance, locally weighted mean phase angle, and phase-preserved denoised image.
• Implemented a three-step model: preprocessing, main binarization using phase features, and postprocessing with adaptive filters.
• Developed a ground truth generation tool, PhaseGT, for simplified and accelerated annotation.

Main Results:

• The proposed phase-based model demonstrated robust performance across multiple benchmark datasets (DIBCO, H-DIBCO, PHIBD, BICKLEY DIARY).
• The postprocessing method successfully improved the performance of other binarization techniques, particularly those with high recall.
• PhaseGT facilitated efficient and accurate ground truth creation for ancient document images.

Conclusions:

• The phase-based binarization approach offers a significant advancement in processing degraded historical documents.
• The developed postprocessing method and PhaseGT tool provide valuable contributions to the field of document image analysis.
• The model's robustness suggests its applicability to a wide range of challenging ancient document imaging scenarios.