Inter - Word Constraints in Visual

A mcthod of using knowledgc about constrainl.~ be/weell words in a technique for reading a running tcxt is presented. This knowledgc IS reprC!itnted as a set of allowablc transitions (c.alled ~inter-word constraint.<;") and a method is given for incorporating this knowledge rcpresentation in a system for rcading visual imag~ of text. This system first anal,vses the shape of a word image to sugg~t a group or neighborhood of words in a dictionary (list of words) that contains an input word. The inter-word constraints are then used to reduct the sil.e of the neighborhood and the smaller neighborhood IS used to 'direct further detailed analysi~ of the input. Thi~ Pf('i\.'C!iS results in a match of the input image to llOe of the words 10 the neighborhood. R~ult.<; are reported in thi'l paper on Ih~ performance and COSt of two reprC!itntations lor inter-word ("nstrain!..'i. The potential of thC!it kno\\. ledgc S()urc~ tt> redu(c the neighborhood Sil.e is explored in a seTies of statislIC<l1 c\perimcnt'i. It is shown that the average size of a nCighborhood em;uulltcred when a tC);t of 150,(XlO words is -readon a word by word halois is redu~ from 16 to about 2. It is also shown empiric.ally that the memory' rieeded for both knowledge representations grows linearly with dictionary ,size and the total additional memory requirement is about twice that needed fOT the original dictionary •.