Segmentation of scanned handwritten text into word fragments

Abstract

Currently, one of the main trends in document flow is its implementation in electronic form, which allows the widespread use of computer technology. Quite often, stored documents are images of scanned texts that are either completely handwritten or partially (for example, signatures under the texts). In this case, it is the handwritten part that is often of the greatest interest. An example is the analysis of the frequency of the use of certain words by the author of the manuscript, the task of detecting falsification of signatures, etc. It is important to note that in most cases we are talking about the identification of word fragments based on their assignment to a class defined by some pattern (precedent). In this paper, one of the stages of solving such a problem is considered, within the framework of which fragments of images that do not contain handwritten characters are determined. In this case, a fragment of a string that obviously does not contain handwritten characters serves as a precedent. It is shown that the analysis of manuscripts in this case is advisable to carry out within the framework of subband representations. The basic relations for the decision function are obtained and a training procedure is proposed that makes it possible to construct a critical region based on a given probability of errors of the first kind.