Performance evaluation and benchmarking of six-page segmentation algorithms

Area of Science:

• Computer Science
• Image Processing
• Pattern Recognition

Background:

• Page segmentation is a critical step in Optical Character Recognition (OCR) systems, often limiting overall performance.
• Existing evaluation metrics for page segmentation are inadequate for diagnosing specific error types and can miss significant segmentation flaws.

Purpose of the Study:

• To introduce a novel vectorial evaluation score for page segmentation in OCR.
• To address the limitations of current scores in identifying over-segmentation, under-segmentation, and mis-segmentation errors.
• To provide a method for pixel-accurate evaluation on arbitrary region shapes with canonical ground truth.

Main Methods:

• Development of a vectorial evaluation score sensitive to key segmentation errors.
• Establishment of a canonical representation for ground truth data.
• Pixel-accurate evaluation of segmentation algorithms on diverse region shapes.

Main Results:

• The proposed vectorial score effectively identifies over-, under-, and mis-segmentation errors.
• The evaluation scheme revealed specific flaws in widely used segmentation algorithms like x-y cut, smearing, and whitespace analysis.
• Demonstrated the superiority of the new scheme in diagnosing segmentation method weaknesses on the UW-III database.

Conclusions:

• The new vectorial score offers a more informative and accurate method for evaluating OCR page segmentation.
• This enhanced evaluation facilitates targeted improvements in segmentation algorithms.
• The method provides a robust framework for benchmarking and understanding segmentation performance.