Keyword selection from word recognition results using definitional overlap

A method is presented to locate a set of potential keywords for a given document in the output of a word recognition algorithm. A clustering step locates words of significant length that occur several times. A word recognition algorithm is applied to these words to generate groups of visually similar alternatives for each image. A simulated annealing algorithm is then used to choose a st of keywords that contains at most one representative from each neighborhood such that an inter-word compatibility measurement is maximized. The compatibility measure is based on the similarity of subject and the definitional overlap of two words as measured by a dictionary. Experimental results are presented that illustrate the ability of the technique to operate in the presence of noise.