A scale space approach for automatically segmenting words from historical handwritten documents

Area of Science:

• Document Image Analysis
• Computer Vision
• Digital Humanities

Background:

• Libraries and museums hold vast handwritten historical documents.
• Automatic word segmentation is crucial for recognition and retrieval.
• Existing methods struggle with noisy historical documents.

Purpose of the Study:

• To develop a novel algorithm for automatic word segmentation of historical handwritten documents.
• To address challenges posed by noise and variability in historical manuscripts.
• To improve the accessibility of historical document collections.

Main Methods:

• A scale space algorithm is employed for word segmentation.
• Page cleaning, line detection via projection profiles, and anisotropic Laplacian filtering are used.
• Optimal scale selection is achieved by maximizing blob extent, followed by bounding box recovery and postprocessing.

Main Results:

• The novel algorithm successfully segments words in noisy historical documents.
• Tested on the George Washington corpus, it achieved a 17% total error rate.
• The method outperformed the state-of-the-art gap metrics algorithm on this dataset.

Conclusions:

• The proposed scale space algorithm offers an effective solution for segmenting historical handwritten documents.
• This technique enhances the potential for automated analysis and retrieval of historical texts.
• The algorithm demonstrates superior performance on challenging real-world historical document images.