Optical Process and Analysis of Historical Documents

The collections of historical books are an important source of information, both for the history of previous periods and for the development of the cultural documentation itself. Although to date, there have been made several attempts of digitalization and electronic navigation, there is not an appropriate frame of optical process and analysis of the content of these collections, consequently a large number of historical books have not been studied yet and remain unexploited. In this thesis, we studied the preprocessing stages which are performed before the recognition process and we focused on the enhancement and segmentation of historical documents. Preprocessing stages play an important role in document image processing since they affect the performance of subsequent processing, such as optical character recognition. At the enhancement stage, we focused on the border removal as well as on the dewarping of document images, which are common problems associated with historical documents. Two methodologies that detect and remove black borders as well as noisy text regions are proposed. Furthermore, optimal page frames of double page document images are detected. The experimental results on several historical documents demonstrate the effectiveness of the proposed techniques. Concerning the warping problem, a coarse-to-fine rectification methodology to compensate for undesirable document image distortions is proposed. To verify the validity of the proposed methodology, experiments have been carried out using indirect evaluation techniques as well as a novel semi-automatic evaluation methodology. At the document image segmentation stage we proposed a novel combination method of complementary text line segmentation techniques. Furthermore, a methodology for character segmentation in historical documents is suggested. Comparative experiments using several historical documents from different languages and time periods prove the efficiency of the proposed technique. Finally, in order to ease the construction of document image segmentation ground-truth that includes text-image alignment we presented an efficient technique.