Reading handwritten words using hierarchical relaxation

Abstract Handwritten words are read (recognized) by applying hierarchical relaxation labeling procedures to a hierarchical description of the word. The H and w ritten W ord R eading S ystem (HWWRS) builds a hierarchical description of the word it reads. HWWRS's first task is to translate the array representation (image) of a word into a geometric graph representation that models the line structures in the word. A stroke graph containing the possible strokes in the word is created by matching structural descriptions of strokes to the geometric graph. The stroke graph is then augmented to a letter-stroke graph that represents the possible letters of the word by matching structural descriptions of letters to the stroke graph. The letter-stroke graph represents the ambiguous segmentations of the word into letters and strokes that possibly overlap one another. It constitutes the hierarchical description of the input word. The hierarchical description is contextually disambiguated and thus simplified by the application of relaxation procedures. This simplified hierarchical description defines one or possibly a few words. HWWRS reads the input word by picking the most likely word from a list of possible words. The system is shown to function correctly for several example words, and experiments involving different relaxation formulas are described.