An approach to estimation of text recognition algorithm parameters 
(character templates) is proposed, based upon the learning sample that 
consists of samples of text line images and corresponding character sequences 
and doesn't require segmentations of these images into images of separate characters. 
Also a method is proposed for template size estimation along with the estimation of its value. 
Formulation of the general learning problem in such a way allows us to simplify 
the process of template estimation for a given text image source significantly 
by removing the need to estimate manually template sizes 
and mark character regions on each image, and leaving only the need to type 
in the text that corresponds to learning images from the sample.