ICR Deciphers Cursive ... and Decodes Handprint

By Arthur Gingrande

A few years ago, computers had as much difficulty reading handwriting as we would with hieroglyphics. If you wanted to feed handwritten information into a computer, you had to laboriously key it in. The automated alternative, intelligent character recognition (ICR) of hand print data, existed but didn’t provide practical accuracy rates. This has all changed, and the mystery of deciphering hand-printed and even cursive writing has been largely solved. ICR has been gee-whiz technology since its introduction to imaging in the late 1980s. Because human hand print is so varied, ICR is complex and uses complicated artificial intelligence algorithms to do its work.

While early generations had relatively low accuracy rates, hand print recognition accuracy has improved significantly. Today, if an end user is willing to use “ICR-friendly” forms that are printed in drop-out ink so that hand print characters are automatically extracted when scanned, chances are that ICR will equal or exceed human accuracy. In practical terms, this can replace 60% of the data entry operators previously required. The labor savings can be enormous.

A few of today’s ICR engines — A2IA, Ceresoft and Parascript — can even accurately recognize cursive words on forms and in other paper environments at industrial speeds. The most prevalent use of cursive handwriting recognition is in check processing environments such as wholesale lockbox, remittance processing and proof-of-deposit applications. By recognizing the handwritten legal amount on the check and using the result to double-check the recognition results of the numeric or courtesy amount on the same check, this comparative procedure can boost check ICR accuracy by as much as 20%.

ICR toolkits have developed an impressive array of features. Standard image processing features include image cleanup, registration, deskew, line removal and enhancement for dot matrix characters and fax images. They offer a wide variety of context analysis procedures and scripting languages for data validation rules. Speed has improved from three characters per second (CPS) back in the late ’80s to more than 1,000 CPS today.

Most ICR engines can recognize multiple languages. Some can recognize Scandinavian languages, Arabic and Kanji. Others do more than recognize hand print; they recognize special math and engineering symbols, check boxes and bar codes. Any ICR toolkit vendor can create a custom “memory” tailored to recognize specific fonts or character sets — for the right price.

Improving Accuracy

High per-character accuracy does not ensure ICR success. In order to do “useful work,” an ICR application must be able to recognize data accurately at the field level. Accurately recognizing 9 out of 10 digits in a telephone number may get you 90% accuracy, but it’s also a wrong number.

Field level accuracy can be boosted significantly by using context analysis, look-up tables, check digits and other data validation tools that can compare raw ICR results with predefined expectations. An example would be checking recognition results of hand-printed addresses against ZIP codes and vice-versa. By comparing the two data sets, multiple-character and multiple-field errors in addresses can be corrected automatically to 100% accuracy without human intervention.

With today’s extraordinary CPU horsepower, mammoth disk space and the availability of cheap memory, enormous amounts of context analysis and data validation routines can be embedded in an ICR engine to improve accuracy. Many analysts argue that incremental differences in raw recognition accuracy no longer matter because context analysis and validation routines can eliminate the initial ICR errors. This is especially true in the world of forms processing, where context is rich and data fields are tightly interrelated.

ICR vendors would be loath to admit that there is little difference in recognition accuracy between products. Jim Tauber from CGK (www.cgk.de) says that accuracy is an obsession at his company. By employing a voting algorithm that analyzes results from the CGK and AEG hand print recognition engines, CGK boosts raw ICR accuracy 20%, he says.

At Mitek (www.miteksys.com), Bridget Baker agrees that accuracy is important. “Mitek has made its reputation recognizing numerics with high accuracy, especially courtesy amounts on checks,” she says.

Likewise, Jacques Germans of reRecognition (www.rerecognition.com), which supplies Datacap, Cardiff and Caere, says his company spends most of its development time enhancing accuracy.

At NCS (www.ncs.com), marketing director Greg Becker says future development will focus on improving raw recognition accuracy rather than adding more forms processing features, such as deskew, noise removal and output to HTML.

“NestorReader has focused almost exclusively on improving raw recognition accuracy by reducing substitution errors and increasing confidence values of the characters we read correctly,” he says.

Not all software developers agree on the importance of engine accuracy. “Raw ICR accuracy really counts when you’re recognizing isolated, hand-printed numbers in the courtesy amount fields on checks,” says Jim Cowen, CEO of Aquracy.com, a Chicago-based remittance processing vendor that uses check recognition. “Apart from comparing the legal and courtesy amounts, checks don’t offer much in the way of context for validation, especially when it comes to proof of deposit applications where you can’t compare ICR results with the OCR of a machine-printed invoice amount.”

An opposing view is offered by Scott Blau, CEO of Datacap, a capture vendor that uses a variety of OCR/ICR engines. “In the real world, the idiosyncrasies of the application and how the integrator copes with them are the factors that invariably determine the throughput of an ICR installation,” Blau says. “I can take any ICR engine and tweak the rules and validation procedures in the application to make it look like the best recognition software in the world, but you also have to give me control over the design of the form.”

Michael Berman, VP of the IDP Companies, a Sudbury, MA, integrator that specializes in document management applications, agrees with Blau. “The bottom line in ICR accuracy is making sure you’ve got the rules and lookup tables right,” he says. “Apart from pure numeric applications, we find little difference between one ICR engine and another.”

If there is a consensus, it is that the less context validation is built in, the more raw recognition accuracy will count.

Taking on Cursive

Thanks to the massive CPU horsepower available today, ICR innovators now have the speed and memory they need to run the extremely complex algorithms required to accurately recognize unconstrained, cursive words and sentences.

Unlike conventional ICR, cursive handwriting recognition does not rely exclusively on classifying individual characters. These engines recognize generic “elements” of whole words, relying heavily on application-specific dictionaries. They also employ ascender/descender analysis, meaning that they consider the pattern of loops above and below the baseline of a word. The longer the word, the greater the odds the correct word will be matched.

By combining recognition of handwriting morphological “elements,” classification of individual characters with ascender/descender analysis and application-specific dictionaries, ICR engines can accurately recognize cursive words. The “killer app” for this is check recognition. By independently recognizing the numeric (courtesy) amount on a check and interactively comparing it with the handwritten (legal) amount, check recognition accuracy can be boosted tremendously over courtesy amount recognition (CAR) alone.

Joe Gregory from Orbograph (www.orbograph.com), Denny Chrismer from Parascript (www.parascript.com) and Bob Foley, sales and marketing director at Ceresoft (www.ceresoft.com), report accuracy gains in the vicinity of 20% when a hand print recognition engine is combined with a cursive recognition engine. Given that there are approximately 160 billion checks processed each year, use of this technology is bound to grow.

Arthur Gingrande is a partner at Imerge Consulting. He can be reached at arthur@imergeconsult.com.

A2iA This company’s Intercheque is a production check processing subsystem that utilizes “A2iA-Binarization.” This is a library of functions that enable the use of grayscale scanners, and it works with A2iA-CheckReader, their flagship ICR product. You input a grayscale image for A2iA-Binarization and it outputs a bitonal image suitable for ICR or screen display. It also features image background removal. A2iA-CheckReader is a recognition application for large batches of checks and remittance tickets that can use both courtesy and legal amount recognition. While CAR (courtesy amount recognition) is used 100% of the time, A2iA software can adjust the percentage of LAR (legal amount recognition) involved. Although LAR demands more computing power, recognition rates improve and the percentage of confusion errors declines. CGK/AEG Computer Gesellschaft Konstanz (CGK) is a subsidiary of Siemens that owns both the CGKand AEG recognition engines, which comprise the largest installed base of ICR engines in the world. Major customers include the US Postal Service, which uses AEG to recognize address blocks on 14 million letters per day. The architectures of AEG and CGK are similar, but CGK is better at character segmentation. The AEG engine is particularly accurate in reading numeric hand print as well as machine print characters — especially degraded fonts and faded dot matrix print. Although each engine is being marketed separately by CGK, there are plans to include the two in a product called Superstar, which will use them in conjunction with a voting algorithm to optimize ICR accuracy. Company tests indicate that the combination increases accuracy 20%. Caere Caere’s Developer’s Kit 2000 provides a single source for multiple recognition technologies including OCR, ICR, barcode, OCR-A, OCR-B, E-13B (MICR) and OMR (Checkbox). The open, 32-bit architecture lets you integrate non-Caere capture and recognition products. The toolkit offers a choice of two ICR recognition engines. The Recognita engine supports numbers only (at 150+ characters per second) and is designed for applications where a restricted character set is desired. The second engine is Kadmos from reRecognition. It’s designed for alphanumeric hand print in a variety of languages. Caere’s toolkit includes an image input subsystem for scanning or importing images; an image preprocessing subsystem to improve image quality; a subsystem that supports multiple recognition technologies; and an export subsystem to format the output. CereSoft CereSoft’s CerebralNet is the basis of their FreeStyle ICR engine, which can recognize machine print and freestyle (cursive) handwriting. The FreeStyle Developers Tool Kit is a modified version of the out-of-the-box FormAgent System. It includes an API and form agent design, template making, form ID, registration and data extraction station modules. FreeStyle can compile multiple engines internally to achieve a throughput of more than 500 characters per second. Gentriqs Gentriqs is the US distributor of the Cleqs ICR engine, which used to be distributed by Gator Bait Software. Cleqs, which means “inkspot” in German, is pattern recognition technology that can accurately classify machine print, MICR, barcode, OCR A & B and optical marks. In addition to English, Cleqs offers language classifiers for most European languages as well as Russian (Cyrillic). It can also learn other fonts and languages within a few weeks. New features of Cleqs version 2.5 include a new preprocessing engine, improved accuracy of ICR and OCR, interpretation of Euro symbols, MICR recognition, improved auto-detection of hand print versus machine-print, built-in setup dialogs, database lookups and new barcode types. In addition to toolkit costs, run-time licenses start at $280 for four characters per sec. up to $30,000 for 300 CPS. Mitek Mitek offers the Quickstrokes ICR engine. QuickStrokes is available as a software-only toolkit or in conjunction with RISC-based, high-speed coprocessor boards. QuickStrokes’ ICR technology can be used to develop a forms processing application from scratch, or it can be integrated into an existing system. The QuickStrokes ICR engine recognizes constrained and unconstrained characters. It separates touching characters and automatically adjusts the character boxes on forms for stray data outside the box. QuickStrokes is known for outstanding recognition of numerical data on forms and checks, and it enjoys an enormous installed base of courtesy amount recognition (CAR) engines among banks. Mitek also resells Parascript’s CheckScript engine, which handles CAR and LAR on checks. NCS NCS’ NestorReader 5.0 offers a set of development tools for building ICR and forms processing applications. The engine accurately interprets handprint, machine print and mark sense data. Forms processing features include Form ID, registration, deskewing and template removal. Form ID allows sorting of forms in a batch by identifying graphical objects or character strings. NestorReader 5.0 registration and deskew features automatically align and re-size images to their original dimensions, producing more accurate and consistent recognition. NCS has focused on increasing accuracy rather than adding forms processing features. Orbograph Orbograph offers OrboCAR, a high-performance courtesy amount recognition (CAR) system that reads diverse check and remittance documents. OrboCAR can be implemented as a turnkey solution or provided as a developer’s toolkit with full API and DLL support. Orbograph also offers OrboCAR Gemini, which combines OrboCAR and Lucent’s LCAR ICRengine. By combining the engines and using intelligent voting schemes, OroboCAR Gemini is said to deliver more accurate recognition results. OrboCar Trinity is Orbograph’s latest recognition product, and it adds legal amount recognition (LAR) to the other two engines. Algorithms automatically cross-validate legal amounts and courtesy amounts. This resolves ambiguity in courtesy amounts, increases the overall read rate, lowers the substitution error rate and reduces the chance of miscoded checks. Parascript Parascript’s Natural Handwriting Recognition (NHR) technology uses a combination of word recognition algorithms and dictionary-driven data validation routines to accurately recognize cursive or natural human handwritten words and phrases. This capability has led to large postal contracts for Parascript using its customized AddressScript application, which recognizes address blocks. Parascript makes the NHR ICR engine available through two SDKs. The FieldScript SDK provides ICR of hand print and cursive data fields for developing forms processing applications. It can read the writer’s natural style, free from constraining formats such as boxes or combs. The CheckScript SDK recognizes and cross-validates both the legal (alpha-numeric) and courtesy (numeric) fields found on a check. ReRecognition Tech ReRecognition Technology manufactures the Kadmos ICR engine and claims 40,000 installed licenses worldwide. The latest version, Kadmos-CL, works with common compilers and functions as a high-speed DLL or library for Windows NT/98, OS/2 and Linux. Kadmos-CL uses a newly developed intelligent segmenting process and is available in all Latin languages. Kadmos recognizes hand print, machine print, mark sense and courtesy amounts, plus norm fonts for OCRA, OCRB, F7B, CMC7, and E13B. Norm fonts are on the bottom of every check, every flight ticket and, in Europe, every form of money transfer. The Kadmos ICR engine is used in the US by Kofax, Datacap, Cardiff, Readsoft and Caere. The license fee for a single copy of Kadmos-CL that runs at 1,000 CPS is priced at around $3,500.

The Major ICR Vendors
Company	Engine name	Product Type	Entry- Level Price	Uncon- strained	Back- ground/ Form removal	Cursive	CAR*	Voting
A2iA Paris, France +33-1-56-53-64-00 www.a2ia.com ProductInfo 220	A2iA- CheckReader	Custom check ICR system	NA	Yes	Yes	LAR**	Yes	Yes
Caere Los Gatos, CA 408-395-5148 www.caere.com ProductInfo 221	Developer Kit 2000	SDK	$5,495	No	Yes	No	No	Yes
Ceresoft Silver Spring, MD 301-445-8413 www.ceresoft.com ProductInfo 222	FormAgent FreeStyle Dev. Toolkit	SDK/ Forms App	$3,000	Yes	No	Yes	Yes	Yes
Gentriqs Inc. Tampa, FL 877-436-8700 www.cleqs.com ProductInfo 223	Cleqs	SDK	$1,238*	Yes	Yes	No	No	No
OrboGraph Bellerica, MA 508-667-6037 www.orbograph.com ProductInfo 224	OrboCar	SDK	$20,000	Yes	No	LAR**	Yes	No
OrboGraph Bellerica, MA 508-667-6037 www.orbograph.com ProductInfo 224	Gemini	SDK	$30,000	Yes	Yes	LAR**	Yes	Yes
OrboGraph Bellerica, MA 508-667-6037 www.orbograph.com ProductInfo 224	Trinity	SDK	$50,000	Yes	Yes	LAR**	No	No
ReRecognition Bayside, NY 718-224-8477 E-mail: jgermans@ rerecognition.com ProductInfo 225	Kadmos-CL	SDK	$3,500	Yes	No	No	No	No
Mitek San Diego, CA 858-635-5900 www.miteksys.com ProductInfo 226	Quickstrokes	SDK/ Forms App	$4,800	Yes	Yes	No	No	No
Mitek San Diego, CA 858-635-5900 www.miteksys.com ProductInfo 226	Quickstrokes (CAR)	SDK/ Forms App	$17,000	Yes	Yes	No	Yes	No
NCS Eden Prairie, MN 800-431-1420 www.ncs.com ProductInfo 227	NestorReader	SDK	$3,500	Yes	Yes	No	No	No
Parascript Niwot, CO 303-381-3100 www.parascript.com ProductInfo 228	CheckScript/ Fieldscript	SDK/ Forms App	$15,000	Yes	No	Yes	Yes	Yes
Siemens CGK/AEG Reston, VA 703-262-2000 www.cgk.de ProductInfo 229	AEG Recognition	SDK	$4,995	Yes	Yes	No	No	No
Siemens CGK/AEG Reston, VA 703-262-2000 www.cgk.de ProductInfo 229	RecoStar	SDK	$4,995	Yes	Yes	No	Yes	No
* Courtesy Amount Recognition      ** Legal Amount Recognition