Digitalización y reconocimiento de documentos manuscritos para la preservación de patrimonio cultural
Abstract
The handwritten manusctipt recognizing process belongs to the iniciatives which lean to cultural patrimony preservation shielded in Libraries and files where there exists a big wealth in documents and even handritten cards that accompany incunable books. This work is point to begin with a research and development proyect oriented to digitalization and recognition of manuscipt materials and the paper presented here discuss diferent algorithms used in the first stage dedicated to noise-clean of the image in order to improve it before the character recognition process begins. Since PrEBi-SeDiCI belong to a network of libraries that interchange digitalized documents by scanning, this document has brought up an extra use related to improvement of the images of interchange documents which presented common problems in its digitalization, such as: borders, impurity, not-centered texts, etc... Although it is not the final purpose of this research, it is still a very usefull skill within the framework of libraries consortium interchange. In order to make the handwritten-text recognition and image digitalization process eficient, it must be preceded by a preprocessing stage of the image to be trated which includes thresholding, noise cleaning, thinning, base-line alignment and image segmentation, among others. Each one of these steps will allow to reduce the injurious variability when recognizing manuscripts texts (noise, random gray levels, slanted chacarters, ink level in different zones), and so increasing the probability of obtaining a suitable text recognition. In this paper, two image thinning methods are considered, implemented and finally an evaluation is carried out obtaining many conclusions related to eficience, speed and requirements, as well as ideas for future implementations. In the first part of the document, some definitions are presented related to the used methods, then the obtenied results are shown over the same set of images applying the proposed theories and finally, some ideas about how to optimized the chosen algorithms are exposed.
