Recognition of cursive, discrete and mixed handwritten words using character, lexical and spatial constraints
John T. Favata · 1993
This research formulates a computational model for off-line handwritten word recognition, where a word can be any mixture of discrete characters, cursive components or touching discrete characters. A hypothesis generation and ranking (HG scR) paradigm is used as the basis of a solution to the general word recognition problem. Three levels of constraints: character, lexical and spatial, are utilized to hypothesize and rank word interpretations. The first and most general of the constraints are the whole character shapes embedded among the word strokes. An orderly search of the word for embedded characters produces many possible word interpretations. The number of interpretations are reduced by applying the lexical level of constraint, which consists of the character-to-character transition rules of the language. The remaining interpretations are ranked and reduced using inter-character spatial relationships that pertain to the relative character size. Surviving interpretations are compared to the lexicon using string matching so as to determine word identity. The general word recognition problem is decomposed into three subproblems, discrete, purely cursive and touching discrete, each with a HG scR recognition algorithm. Words with discrete characters are naturally segmented into isolated characters and recognized. Purely cursive words are recognized by generating a series of segmentation points on the word and searching for characters among the segmentation points using a special character recognizer. Plausible interpretations are generated and ranked using lexical and spatial constraints. Words with touching discrete characters are analyzed by making a series of word length hypotheses and segmenting the word under these hypotheses. Hypotheses that yield high-confidence characters are retained as possible interpretations of the word. A general control structure combines these three modules into a general HG scR system which can recognize any form of handwritten word. The control structure tries to determine the type of each word component and activates the appropriate HG scR recognition module. In those cases where a clear determination cannot be made, several HG scR modules will be activated to analyze the component.