Model-Directed Document Image Analysis

If current OCR engineering trends continue, then, we believe, \general{purpose" systems | that is, fully automatic and nonretargetable systems | will leave many potential users unsat-issed, and lucrative application niches unnlled, for years to come. However, for users who care enough to volunteer some manual eeort | to help customize the system to their document(s) | signiicantly higher accuracy may be achievable, without delay. We discuss in detail two state{ of{the{art document recognition systems | Lu-cent Technologies' Table Reader System (TRS) and Xerox's \document image decoding" (DID) research prototype | which yield high accuracy by reliance on explicitly stated models of properties of the target document, whether iconic (known typefaces and image degradations), geometric (restricted classes of layouts), or symbolic (linguistic and pragmatic contextual constraints). How great are the performance advantages that can be realized by sacriicing automation in these ways? To what extent can the necessary customizations be (semi{)automated? We outline recent and planned research at Xerox PARC motivated by these questions. The dominant type of present{day commercial OCR system, whether on the desktop or in service{bureau settings, is designed to operate fully automatically, refusing to accept guidance 0 Invited unrefereed talk presented at the DOD-from the user. The majority of desk{top users welcome this since they are untrained and impatient with inconvenience. There is a similar reliance on more or less completely automatic operation in almost all of the highly specialized OCR application niches such as postal{code and nancial{document processing, even though their costly equipment is tended by trained staa in controlled service{bureau settings. In this case, it is largely the daunting throughput requirements that dictate fully automatic operation. Both of these user communities | the casual SOHO users and the sophisticated special-document users | tolerate surprisingly low performance. The latest competitive studies, at UNLV in 1996 1], showed, for example, that desk-top OCR packages misrecognize 3{15% of characters | an intolerably high error rate, most users would agree | in over 40% of magazine pages: for other document categories, performance was far worse. The best current systems for reading hand{written courtesy amounts on checks 2] are tuned to reject 33{55% of the input in order to hold substitution errors below 1%. Similarly, the best handwritten postal-address readers fail to \\-nalize" 35% of the input 3]. All of these technologies are improvable, of course, and are improving: but slowly and at a high cost. The UNLV data suggested …