Difficulty in Predicting OCR Accuracy

The task of predicting OCR accuracy is very complex. Furthermore, to the best of the author's knowledge, no previous work has been done in this area; therefore, no reference titles can be given.

OCR algorithms seem to be affected by a myriad of different problems. However, the following three general ``problem groups'' can be identified:

• Typographical problems. Unusual fonts or complex typesetting often cause problems for OCR algorithms. Figures 1.4 and 1.5 shows some examples of complex images and their resulting OCR output.
  

  

  The Title Index lists programs and playdates by network,

  so you can consider only those channels you get in your home.

  tarn identifies Pimmiems of programs by network.

  U identifies closed Capuoned programs for the hearing impaired.

  * indica a un programa que se puede recibir en Espahol donde disponible.

  (Indicates programs that can be received in Spanish, where available.

  * identifies films of Superior qudity~

  * identifies films which are ~adfrfo~tv or made-forccable prewieres.

  Figure 1.4: Typographical Problems for OCR Algorithms (1/2)

  

  

  

  OLOR US BUSY THIS MONTH: Nk7E'VE FOUND A GREAT

  bike-and-wine trip in Italy a superb gaidebook to go with

  it, a spirit from Australia and a crisp, delightful white wine

  made at an estate with a seventeenth-century pAace overlooking the

  Rhine. That's a lot of territory to cover on one page. Have a look.

  TheToplO

  Mete air your best betsfot. die uionth, selected fr die Ban Apperit

  Tasting Panel, wilicli aleets weekI, under die dijeetion of wine and

  spitits edit()r Antliony Dicis Blite and his associate. Jack R. `Veiner.

  1990 Parducci Wane Cellars, Johannisberg Riesling. North Coast ($6).

  Snappy `vith lively acidit,, and fine apple and peach nuances.

  1992 Jacob's Creek, Chardonnay, South Eastern Australia ($8).

  A charming white that's crisp and fively "4th great clean fruit.

  1991 Prosper Maufoux, C6tes du Rh\&ne, France ($8). A dense red ~irie

  featuring "leathery., black cherry and pepprry fruit and a soft finish,"

  says panel member Peter Kay of The Stouffer Stanford Court hotel.

  Figure 1.5: Typographical Problems for OCR Algorithms (2/2)

• Linguistic problems. As stated in [5], linguistic content can also be problematic for OCR algorithms. OCR systems use lexicons to solve recognition ambiguity. This greatly improves OCR accuracy, but can also be a drawback when proper names, acronyms, or other words not likely to be in the lexicon are part of the text to be recognized. In these cases, the number of lexicon-dependent corrections are reduced to a minimum. Furthermore, an exaggerated sensitivity to the lexicon could be the cause for the introduction of new errors (i.e., an OCR device with an exaggerated sensitivity may modify an otherwise correctly recognized word to force it to match the system's lexicon.)

• Image problems. In the OCR community, it is well known that image defects directly affect the performance of OCR algorithms. Furthermore, image problems account for the majority of OCR errors [10].

Image defects constitute the bulk of the problems associated with OCR algorithms (see Chapter 2 of this thesis). Therefore, the focus of this work is on the detection of image problems. By better understanding image defects and subsequently implementing OCR algorithms that are sensitive to these type of problems, it could be possible to achieve acceptable accuracy ranges (95%-98%, [4]) for most printed pages. To achieve near perfect (99.5%-100%) recognition, however, typographical as well as linguistic problems would have to be addressed.