Document

In this paper we will present a set of experiments using large digitalized collections of books to show that logical structures can be extracted with a good quality when working at document level. The proposed solution relies on a twofold method: first specific logical elements are recognized by a given method. Then models for the recognized elements are generated by combining layout, content and labeling information. Model inference is made possible at document level, a level which promotes frequent occurrences of document structures. These inferred models combining several kinds of information are used to correct noisy data, typically zoning, OCR and labeling errors produced by previous processing steps. This method is illustrated by the detection of two document structures: page numbers and chapter headings, two navigating elements required by digital libraries.