About precedent identification of image fragments scanned manuscript

At present, large repositories of data obtained by scanning handwritten texts have been accumulated. A significant part of them are presented by scanned printed documents, which contain handwritten signatures of officials. The images of texts obtained in the process of scanning are often subjected to computer analysis in connection with one or another need. Search for fragments of these images, containing preset word forms, for example, in philology when studying the frequency of use of certain words by the same author is of significant interest. You can also indicate cases of word search from the standpoint of ensuring the safety of socio-economic processes. An important example is the detection of falsification of signatures of officials, etc. A feature of the automatic search for identical word fragments in images of scanned documents is the ability to identify them using only one text sample (precedent), which requires the creation of a special machine learning technique. In the presented article a decisive procedure for classifying word fragments of images of scanned handwritten text as identical to a given precedent has been developed. It was proposed to use the projection of vectors onto the eigenvectors of subband matrices corresponding to nonzero eigenvalues as elements of the feature space. A method for the formation of total subband matrices is substantiated on the basis of the introduced concept of information subbands in the area of spatial frequencies. A training procedure based on one precedent is proposed. This procedure is based on the developed method for generating vectors, the totality of which simulates the training sample. An algorithm for processing images when searching for identical to a given fragment was formed.