A computer-implemented method comprises generating a description of a character symbol from a binarized image; comparing a template for the character symbol with the description of the character symbol based on a reference description, wherein the template comprises a grid of cells, a set of local features which may be present in the grid of cells, the reference description specifying which member of the set of local features should be present or absent in the grid of cells, and a threshold of an accepted deviation with the description of the character symbol; assigning a penalty value to the description of the character symbol via a cost function when a discrepancy exists based on the comparing; selecting the template as a match candidate for the character symbol when the penalty value is below the threshold; recognizing the character symbol based on the selecting.