Using combination of structural, feature and raster classifiers for recognition of handprinted characters

The paper presents a novel method for off- and on-line handprinted character recognition using structural patterns. A character is described as a set of structural elements such as segments, arcs, circles and points. Possible relative placement of elements is described with the help of spatial relations, which are expressed as fuzzy logic predicates. The structural pattern is matched against a character image by establishing correspondence between structural elements and parts of the image which satisfy all spatial relations. The developed matching procedure can successfully find this correspondence on broken and distorted images. The found match for the structural pattern effectively segments a character into meaningful parts, thus giving access to plenty of information about character structure and properties. This allows one to develop simple yet accurate pairwise discriminant functions for similar characters and use them to further increase recognition accuracy. The described system has been implemented and structural patterns for digits and the Cyrillic alphabet have been developed. Detailed results of experimentation are presented.