Character templates learning for textual images recognition as an example of learning in structural recognition

Document recognition for digital libraries is characterized by high requirements to a recognition quality and processing of significant amount of single-type documents. So this is a perfect area for single-font approaches because they provide a smaller error rate comparing to multifont approaches and a learning of the font is carried out relatively rarely, because of significant amount of single-type documents. Traditionally character templates learning is performed for separated characters on a basis of the set of character examples. It leads to recognition errors like in situations when closely placed parts of neighbouring characters are recognized as a single, separate character. We propose another approach to character templates learning. Namely such templates must be constructed that the result of recognition of a text line image as a whole must match to a text string specified by a teacher. The approach guarantees that not only images of separate characters will be recognized correctly, but also the segmentation of the whole text image into characters will be performed without errors. So in our approach a learning sample consists not from labeled images of separated characters, but from text line images with corresponding text strings