报告摘要:

Document  Image  Analysis  for  Digital  Libraries

The  rapid  growth  of  digital  libraries  (DLs)  worldwide  poses  many challenges  for  document  image  analysis  (DIA)  research  and  development. DLs  promise  to  offer  more  people  access  to  larger  document  collections, and  at  far  greater  speed,  than  physical  libraries  can.  But  DLs  also tend  to  serve  poorly  many  types  of  non-digital  human-legible  media such  as  printed  and  handwritten documents. These  documents,  in  their physical  (undigitized)  form,  are  easy  for  people  to  read  and  browse, whereas  when  they  are  accessed  through  DLs  they  often  lose  these advantages  while  of  course  lacking  many  advantages  of  symbolically encoded  information.  This  talk  explores  these  issues  and  illustrates them  with  case  studies  arising  in  several  DL  projects  in  the  US. Difficult open  DIA technical  problems  in  DL  applications  areidentified,  forexample  during  image  capture,  early  image  processing, content  extraction and  recognition,  image  presentation,  and  retrieval---and  in  personal  and  interactive  DL  settings.    Recent  research at  Lehigh  Univ.  on  highly  versatile  document  image  contentextraction  algorithms  using  fast  hashed  k-D  tree  classifiers  is  alsosummarized. [Joint  work  with  Michael  Moll  and  Matthew  Casey.]

Human  Interactive  Proofs 

Internet  services  offered  for  human  use  are  suffering  abuse  by computer  programs  ('bots,  spiders,  scrapers,  etc).  We  can  defend against  such  attacks  with  CAPTCHAs---Completely  Automatic  Public Turing  tests  to  tell  Computers  and  Human  Apart---which  are  special cases  of  `human  interactive  proofs'  (HIPs),  security  protocols  allowing people  easily  to  authenticate  themselves  over  networks  as  members of  given  groups.   I  will  review  six  years  of  HIP  R&D,  share  highlights of  the  first  two  HIP  workshops  (the  most  recent  held  at  Lehigh Univ.),  and  describe  CAPTCHAs  now  in  use  and  on  the  horizon. One  of  the  best  ways  to  engineer  a  CAPTCHA  is  to  exploit  the  gap in  ability  between  humans  and  machines  in  attempting  to  read  images of  text.  I  will  analyze  the  strengths  and  weaknesses  of  several such  reading-based  CAPTCHAs,  and  give  details  of  ScatterType, developed  here  in  collaboration  with  Avaya  Labs.  Its  legibility  hasbeen  validated  by  experiments  on  human  subjects.  Recently  we  have explored  tradeoffs  between  the  familiarity  of  challenge  strings and  image  degradation  in  an  attempt  to  control  the  difficulty  of CAPTCHA  recognition.

[Joint  work  with  Terry  Riopka,  Michael  Moll,  Dan  Lopresti,  Sui-Yu  Wang, Jon  Bentley,  and  Colin  Mallows.] 

报告人简介:

Dr. Baird is a Professor of Computer Science & Engineering at Lehigh Univ. and (with Dan Lopresti) heads up Lehigh's Pattern Recognition Research lab. Prior to joining academia he was a researcher and research manager at Bell Labs and the Xerox Palo Alto Research Center. He has been elected Fellow of the IEEE and also of the IAPR, and has received an ICDAR Outstanding Contributions award. He has served on the Editorial Board of several journals including IEEE Trans on PAMI and CVIU; and he was a founding member of the Editorial Board of the Int'l J. on Document Analysis and Recognition. He has published three books and seventy-six technical articles, and he holds seven patents. He has been founder, co-organizer, or program co-chair for six conferences and workshops.