COMPUTATIONAL APPROACH TO VISUAL WORD RECOGNITION: HYPOTHESIS GENERATION AND TESTING.

An algorithm for reading images of v.ords of text ,is pre'!lt'nted and analyzed. The algonthm IS ba!led on a model of the human reading process '.A.'hu.:h suggests that word recognl' tlon has two stages: hypothesis generation and hypothesIs test· mg. Given an input word. h~'pothesis generation finds a group of candidate words from a given '\"!Xabulary. 1I~'pothesls testing is a feature testing strategy to discover the word In this group that best matches the '.A.'ord in the input image. This p;ipt'r concentrates on the hypothesis tt'iting ~tage. which is formulated as a tree seilrch problem In which a small number of tests are executed to recognl7.1! an Input '.A.'ord. A statistical study using a text of over 8()O.0(10 words sho'.A.·s that at most th ree tests. from among 212 different tests. '.A.'ould han to be executed to recognize any '.A.ord in that text. A complete analysIs of this method IS alSfl g!\C'n for smaller text. ... E~ pt'ri' ments with word images that show the feaslhiht:; of thiS technique are also described. e,g~ images of 100 '.A.·oras in five different fonts were recognized '.A.'ith 92% to 97<;(accuracy.