Adaptive Methods for Robust Document Image Understanding

A vast amount of digital document material is continuously being produced as part of major digitization efforts around the world. In this context, generic and efficient automatic solutions for document image understanding represent a stringent necessity. We propose a generic framework for document image understanding systems, usable for practically any document types available in digital form. Following the introduced workflow, we shift our attention to each of the following processing stages in turn: quality assurance, image enhancement, color reduction and binarization, skew and orientation detection, page segmentation and logical layout analysis. We review the state of the art in each area, identify current deficiencies, point out promising directions and give specific guidelines for future investigation. We address some of the identified issues by means of novel algorithmic solutions putting special focus on generality, computational efficiency and the exploitation of all available sources of information. More specifically, we introduce the following original methods: a fully automatic detection of color reference targets in digitized material, accurate foreground extraction from color historical documents, font enhancement for hot metal typesetted prints, a theoretically optimal solution for the document binarization problem from both computational complexityand threshold selection point of view, a layout-independent skew and orientation detection, a robust and versatile page segmentation method, a semi-automatic front page detection algorithm and a complete framework for article segmentation in periodical publications. The proposed methods are experimentally evaluated on large datasets consisting of real-life heterogeneous document scans. The obtained results show that a document understanding system combining these modules is able to robustly process a wide variety of documents with good overall accuracy.