Document representation refinement for precise region description

Precise description of layout entities (content regions on a page) is crucial for all but the most trivial document analysis and recognition applications. The output of layout analysis methods and state-of-the-art OCR systems varies significantly, from bounding boxes (e.g. Tesseract) to stacks of text line rectangles (e.g. ABBYY FineReader). There is a clear need for a consistent and accurate representation of regions (e.g. text paragraphs, graphics entities etc.) for further processing, correction and performance evaluation (comparison of segmentation results with ground truth regions). This paper describes a method for refinement of document representations by fitting polygons around lower-level layout objects (such as text lines, words and glyphs) in a systematic way that reconstructs region outlines and preserves the fine details of complex layouts. Experimental results on a standard dataset demonstrate the validity and usefulness of the proposed approach.