Not so long ago OCR was an exotic application found in government computers and large corporations. Since hitting the desktop OCR has become an important part of data entry systems. It won't be long before ordinary desktop computers can read your handwriting.
The market for optical character recognition can be divided between the desktop market and OCR production. OCR production is aimed at corporate users who do high-speed scanning of forms and documents. Production scanning requires speed and accuracy. Even an error rate less than 1% results in thousands of scanning errors
Shrink-wrapped OCR packages use the same recognition engines found in their big brothers. This ensures they provide the same robust recognition and error correction production scanners require. Even SOHO users who scan only a few documents at a time, get advanced artificial intelligence technology.
The existence of the shrink-wrapped versions of these OCR engines makes them easy to review. While this software reviewer can not easily put together a large scale production system, I can easily test shrink-wrapped software. The speed and quality of the OCR engines used in packages aimed at the desktop should suggest how well the same engines will function in a production environment
Caere's (Los Gatos, CA 408-395-7000) OmniPage Pro (full $500, upgrade $130) is probably the best known OCR package available. OmniPage started in the SOHO market. Version 8.0 has a well polished user interface, an easy-to-use manual and plenty of Wizards to smooth over the complexities of optical character recognition
An added feature of OmniPage is the ScanFix for OmniPro ($100-$400) daughter application. ScanFix is an image processing application from TMSSequoia (Stillwater, OK 405-377-0880) that cleans up scans for character recognition. A standalone version of ScanFix is also available for other recognition systems. The company even has a version of ScanFix with a Java-based user interface
Caere's chief rival for desktop OCR is Xerox's (Peabody, MA 978-977-2000) TextBridge Pro ($80). This has been spun off into its own software division called ScanSoft. TextBridge is inexpensive, professional level OCR from a company that prides itself on how well it handles documents - both paper and electronic
WordScan's creator, Calera, was purchased by Caere a couple of years ago. WordScan and OmniPage are very different products. While OmniPage traditionally served the desktop user, WordScan has always served the production scanning market. And while OmniPage is a strong retail product, WordScan is sold primarily through the OEM channel
WordScan was the most forbidding package I reviewed. There's a 500 page manual. The program comes on 13 floppy disks, the only one to eschew CD-ROM. A closer look reveals that the manual is written in five languages. The documentation is really only 100 pages long. WordScan has the most modest hardware requirements. It will run on a 386-based PC
Less well known in the SOHO market is Expervision's (Fremont, CA 510-623-7071) TypeReader ($400, upgrade $150). TypeReader includes several applications in one box. TypeReader Professional for OCR, TypeReader EZ, TypeReader Imaging for document viewing and scanning, and TypeReader Manager to keep track of scanned and recognized files
Finally there's NeuroTalker ($100) from INM (Waterloo, Ontario 519-746-3890). We looked at NeuroTalker at AIIM 97. NeuroTalker has the simplest packaging. The documentation was a small spiral bound book
Because NeuroTalker is commonly integrated into production OCR and document management systems, the company also has a toolkit to ease programming with the NeuroTalker engine
For these tests I put together a suite of documents to scan. There were clearly printed letters in Courier and Times Roman and fonts and pages taken from magazines with colored backgrounds
There were a number of OCR busting faxes with serif and san serif fonts in different point sizes. The faxes were as clearly printed as a fax can get. All scans were performed on a Fujitsu (San Jose, CA 408-432-6333) ScanPartner 600C ($1,800) with an automatic sheet feeder. I used a 200 megahertz Pentium MMX to do the scanning
OCR Test
NeuroTalker's finest OCR moments came when it recognized the clearly printed Courier and Times Roman fonts. It also recognized some of the faxed pages with tolerable error rates. NeuroTalker maintained page formatting to a reasonable extent
NeuroTalker's user interface was the simplest one I tested. It had all the features needed for OCR including zoning tools, auto deskew and rotation. After a document has been recognized you can bring it up in the proofreader. The proofreader lets you tab through the recognition errors while keeping a window on the screen showing the original scan. NeuroTalker did a good job flagging its own errors. The proofreader was easy to use
NeuroTalker was not the fastest OCR package, but it recognized pages in a reasonable amount of time. Documents were scanned at NeuroTalker's recommended 200 dpi. When recognizing scans made at higher resolutions, its accuracy on high-quality documents improved. High-resolution scans took longer to recognize and some poorly printed documents were impossible to recognize
Caere's WordScan has excellent recognition with better quality documents. It gave good results with the faxed documents. It was about as fast as NeuroTalker. WordScan Plus doubled the top inch every scan and it wouldn't scan properly when the Fujitsu scanner's document feeder was engaged. This meant each page had to be scanned individually, a tedious process. Wordscan was the only OCR application with a scanning glitch
While Wordscan Plus had an easy-to-use interface, the learning curve was steeper than other OCR programs. The WordScan Plus OCR engine is used in Caere's M/Series production scanning system
Expervision's TypeReader was the fastest OCR software. Most pages were processed in 10 seconds or less. Many were processed in less than five seconds. It had an excellent user interface. You can manually zone documents in TypeReader, but its built-in automatic zoning was very good. TypeReader did a good job maintaining all of the document formatting
Overall, the most accurate scanning was from Xerox's TextBridge Pro. Its proofreading module was also one of the best. It wasn't the fastest OCR, but most of the OCR time was spent automatically zoning documents. Recognition time could probably be improved with manual zoning
There were a few surprises in how TextBridge maintained document formatting. TextBridge would measure the dimensions of the text zones in the original document. The recognized text was then formatted to fill the same amount of space in the final document. This led to some surprising results. The words in the finished document would be widely spaced to fill the size of the original. Text on the finished document was kept in the same relative positions as the originals. While TextBridge's final product looked strange, it was easy to reformat in a word processor
The best user interface was on Caere's flagship product, OmniPage Pro. Thumbnails of scanned documents were displayed in their own window. This made it easy to navigate around a multi-paged document. Zoning, whether manual or automatic, was clearly displayed and easy to work with. Manually created zones could be saved in a zone template file. This is useful if you regularly scan many documents with the same layout. The proofreader module suggested corrections to doubtful words
Although OmniPage Pro is a slower than TextBridge, OmniPage's automatic zoning is faster. As a result, OmniPage edged out TextBridge on the documents that have easier to read text and more complicated formatting
All the programs in the test have excellent recognition. They all have their strengths, and each excels at recognizing certain document types
In my testing, I found the best ones were TypeReader, TextBridge, and OmniPage. All three of these packages did excellent jobs, even on the toughest documents. All were awarded an Editors' Choice award. Which you choose depends on your needs. For maximum recognition speed you can't beat Typereader. Its user interface is not as good as TextBridge and OmniPage. That makes it a little more difficult to learn and use
TextBridge has fast recognition with relatively slow zoning. TextBridge is the OCR of choice if you scan a lot of poor quality documents with simple formatting. If you scan a lot of good quality documents with lots of graphics, tables and such, OmniPage Pro is the better choice
OCR and Document Management
When OCR was introduced a few years ago most OCR products were standalone products. As OCR has improved it has become mainstream. It's now hard to find a document management system or scanning station without some sort of built-in OCR
Although handwriting recognition (ICR) has improved greatly, it's still early days for this technology. Handwriting recognition is being integrated into more and more scanning and document management systems. Handwritten forms present the most complicated data entry problems. Key-entering information is expensive. Handwriting and handprint recognition reduce the cost of inputting data by automating the process. No OCR/ICR system is perfect. Strategies for dealing with errors in document recognition include manual proofreading and voting. Proofreading requires someone to look at the recognized document and compare it to the original. For many users, the quality of the user interface is important
AFPS Pro ($10,000) from Top Image Systems (Ramat-Gan, Israel 011-972-3-752-5626) has built-in image processing and a recognition voting engine. Once a job is scanned and recognized, a system administrator can route it to one of any number of proofreading stations
The proofreading stations can be set to look at different forms or even a specific part of a form. AFPS Pro proofreading stations are highly customizable. Operators can set up their workstations any way they want. Once an operator has finished a job, it is either routed to the next proofreading station or sent to the next part of the workflow system. AFPS Pro lets the scanning administrator monitor the entire recognition/proofreading process and reroute jobs when bottlenecks show up
Software companies try to get computers to scan and recognize characters as accurately as possible. One way is to have multiple recognition engines vote on the characters. Each OCR engine has strengths and weaknesses. Voting takes a consensus of two to five engines. This can take more time and yield more accuracy
For voting systems to work they require two things. First, the OCR/ICR engines combined by the voting system should be dissimilar. Second, voting systems need advanced preprocessing. One OCR engine may be good at small type, while another is good with italics and script fonts. The voting engine needs to know the requirements of a document. It then weighs the voting in favor of the OCR engine best suited to the text being recognized.
The tradeoff with voting recognition engines is speed. When multiple OCR engines are used, it takes longer for all the engines to have a go at the documents. Even if you have a computer with multiple processors, recognition is only as fast as the slowest OCR engine. Voting engines work best when the volume of scanned documents is high enough to make manual proofreading costly
PrimeOCR ($1,700-$4,500) from Prime Recognition (San Carlos, CA 650-637-8382) has a voting engine that combines the results of up to five OCR engines. PrimeOCR is more expensive than many single OCR engine systems, but it makes up for the added cost in improved accuracy
PrimeOCR saves recognition time with a feature they call Selective Voting. If it looks like a document is easy to process, only one OCR engine is used. There is no need to process a document with multiple recognition engines if one engine will get almost 100% accuracy. Multiple engines are only used on hard to recognize documents
While handwriting recognition engines achieve a remarkable level of accuracy, many still require that letters be carefully printed and kept separate from each other. If you have ever filled out a form with rows of little boxes where you are instructed to print, you have filled out a form designed for ICR
Hand-in-hand with handwriting recognition is forms recognition. Because of the special zoning requirements of handwriting recognition, the computer must know what kind of form it is looking at before it begins the recognition process. Forms recognition can be achieved by printing an identifying mark on the form. These include barcodes, glyphs or letter codes that are read when the form is processed for OCR
By the time you read this, FormWare (Park City, UT 801-645-9600), the result of a recent merger between TextWare and Symbus, will have the new version of InScript ($10,000-$100,000) ready and shipping
Instead of using a code or symbol to identify a form, InScript uses neural network technology similar to that used in handwriting recognition to recognize forms. InScript's forms recognition technology lets forms be freely designed. A handful of scans of a blank form are all that's necessary to train InScript to recognize it
Cardiff's (Vista, CA 760-752-5200) Teleform ($1,500-$5,000) uses glyphs, called cornerstones by Cardiff, to recognize forms. The cornerstones are in the corners of the form. When scanned, the cornerstone yields a histogram of gray levels that are different for each form. Non-Teleform forms can be recognized by selecting a zone on the scan. A histogram can be created for the zone that is unique for each different form
NCS's (Minneapolis, MN 401-334-4811) forms recognition software, Accra ($30,000-$500,000), uses OCR engines by Caere and handprint recognition by Nestor, another NCS product. Accra has a full set of forms recognition features including job workflow and contextual editing. NCS provides complete forms processing solutions. They do everything from building systems to designing and printing forms
Image Access' (Boca Raton, FL 561-995-8334) two OCR products, BScan ($2,500) and UpFront ($2,000) read data off of forms for indexing purposes. The user creates a zone template and BScan converts the text within the selected zones to database fields. UpFront has a menu-driven user-interface. BScan has a customizable interface. Image Access uses Expervision's TypeReader OCR engine and their own proprietary engine. You can train this engine to recognize custom fonts
ReadSoft's (Chicago, IL 312-867-1008) main product is Eyes and Hands for Forms ($9,000). Like a lot of forms recognition products, Eyes and Hands includes OCR and handprint recognition. Eyes and Hands software will process hundreds of pages per minute and is used in production level forms processing for such things as coupons. ReadSoft uses their own proprietary recognition engines
Microsystems' (Tampa, FL 813-222-0414) flagship product, OCR for Forms ($12,000), was one of the first forms recognition systems to include image processing to make forms recognition and processing easier. Since then, the program has added many features and has become modular. Users only buy features they need. OCR for Forms supports data validation and custom workflows. It supports recognition engines from Caere, Xerox, Nestor and GatorBait.
Forms processing, OCR and handwriting recognition software are often incorporated in large document management systems. Dakota Imaging's (Columbia, MD 410-381-3113) Transform ($20,000), formerly known as GroKKer FPS, is a modular software package designed to be used with Dakota's large document management system. It helps companies that want to let workers work at home with a module for remote site forms processing.
Alos' (Montgomery, NY 914-457-4400) DocuWare fully integrates OCR/ICR and is included in the $3,600 purchase price for the basic system. No external module is needed. DocuWare's OCR uses the Maxsoft-OCRON (Fremont, CA 510-445-8610) optical character recognition engine
For the ultimate in production OCR, many users look to hardware solutions. Hardware OCR is best suited to OCR/ICR applications where tens or even hundreds of thousands of documents must be scanned each day. One supplier of such high speed systems is Scan-Optics (Manchester, CT 860-645-7878). Their Series 9000 scanner ($300,000) can OCR and process up to 12,000 letter-size pages or 30,000 check size documents per hour
BancTec's (Dallas, TX 972-341-4000) E-Series ($100,000+) scanners will process up to 1,150 small documents per minute with OCR, MICR and ICR. The S-Series ($45,000-$90,000) scanners provide full character recognition at up to 220 pages per minute. The S-Series is designed to recognize larger forms as well as small documents. Its vacuum transport is designed to accept pages very fast without jamming. The 9500 Workstation ($200,000) is an all in one solution for smaller volume requirements
RecoStar character recognition software from CGK (Konstanz, Germany 011-49 7531-874447) is fast. It has to be. It will be used to recognize the forms for the Year 2000 Census. RecoStar is available by itself or bundled with the company's document processing system or with their scanning hardware.
