Unified Performance Evaluation for OCR Zoning: Calculating Page Segmentation's Score, That Includes Text Zones, Tables and Non-text Objects

摘要

The optical character recognition (OCR) systems decompose printed pages into a set of text zones, tables and nontext objects, such as pictures and charts. This part of OCR process is known as the page zoning task. In the paper we present the methodology for assessing the page zoning as a whole task. Many authors evaluate the locations of tables, pictures, and text separately. The key advantage of the proposed system is that it naturally combines the evaluation for text and tables locations, and it is resistant to most segmentation's ambiguities. We calculate score for texts and tables, basing on ground-truth character locations. The score for non-text objects locations is based on areas matching. These scores are combined to get the final page score.