 |
|
|
|
|
|
|
|
|
|
| |
 |
|
|
|
Contents
Overview
|
JustICR Engine is a Software Development Kit (SDK) for integrating advanced ICR
(handprint) OCR (machine printed) recognition technology into 32-bit Windows
applications.
JustICR Engine is a DLL (Dynamic Link Library) and a training tool. The Engine API
(Application Programming Interface) can be easily used by C/C++, Visual Basic or any
other development tools supporting DLL components.
The engine comes with a complete evaluation kit. Five minutes Demo and you
are ready to do the evaluation. The training tool gives you all what you need in order to
make the evaluation including a Statistics tool. In most cases,
you can do the evaluation of JustICR with your material within several hours
without writing a singe line of code (!) and without opening the user's
manual…
|
Where to use JustICR Engine?
|
JustICR Engine may be used in image or document processing applications that
need accurate and flexible ICR/OCR capabilities. The applications may be the
following:
Processing of machine-readable forms filled in by hand
Archiving and document processing applications
|
Why JustICR?
Easily trainable recognition engine
|
The key feature of JustICR is the ability to train a new font or handwriting
style in a very short period. In most of the form processing applications there
are "problematic" fields that you wish to recognize better than the results that
can be obtained from the "off-the-shelf" recognizers. Below are several such
cases:
| 1. | Low quality handwritten fields: in some cases you have poor quality of the
scanning in handwritten fields. In such cases you may have many broken characters
or too dirty characters. In these cases the recognition results of the best
"off-the-shelf" engines is significantly reduced. If you trained these fields
specifically with our engine you may achieve much better results. The best way,
according to our experience, is to use voting between our engine and other engines
that you use normally. |
| 2. | Low quality machine print fields: there are cases that the printed field
is of poor quality, such as the case of "stamped field" (FedEx forms, Airborne
forms, Credit Card vouchers, flight tickets...). |
| 3. | Sometimes you have a "sensitive field" that is machine print, and you need
to produce 100% recognition rate for this field. Teaching our engine with the
specific font of this field may help you to be close to this target. |
| 4. | For applications outside US, the handwriting style may be different than the
handwriting expected by the "off-the-shelf" engines. We have many examples for
this case (probably every country except US, Germany and UK...). This is the
reason that in many census projects our engine played a significant role to win
the project. In this case, according to our experience, the best results achieved
by using voting methods with other engines. |
| 5. | Some applications have a specific field that contains specific marks. These,
of course, can be trained with our engine too. One example how to use this feature
comes from the educational market in United Kingdom. The application included
tests that were submitted by the students. In these tests the student did have to
put a mark on the correct answer, and there were several types of mark signs. The
ability to teach these mark signs with JustICR solved this problem
immediately… |
|
Dictionaries
|
It is possible to give to JustICR engine dictionaries. The dictionary may be a
large dictionary (such as the full English/Spanish/German... vocabulary, which may
contain 100,000 to 1,000,000 words). The recognition results are improved
significantly even if the dictionary is not full (it does not contain all the
possible words). One example can be from an application in Germany in which one of
the fields was a First Name (hand print). The recognition rate jumped from 89% per
character to more than 97% per character by using a dictionary of first names.
Equivalent results achieved with geographic places, or "descriptive fields" in
which the field contains a phrase that describes something (such as: occupation,
religion, etc.).
|
Looking for a font in a form
|
There are many applications in which there is a big amount of form types
(hundreds); some of them are "variants" of the same form. There are many examples;
one of them can be demonstrated in FedEx forms. You can fine in these forms a
number written in a different font (OCRA in this example). This number stands for
the form type (different variants of the same form have the same number). In such
cases you may train this font, and ask JustICR to find this number anywhere in
this form. This feature is generally useful for very big customers that have many
types of forms.
|
Success Stories
|
Some of the key installations of JustICR engine:
FedEx, USA
Social Security Administration, USA
Deutsche Post, Germany
Swiss Post
European credit card companies:
|
Visa projects: Israel, Portugal, and Croatia
MasterCard - Israel
|
Several census projects:
|
Brazil, India, Italy, Ireland, Kenya, Cyprus
|
DataSel: Turkish Ministry of Statistics; Garanti Bank; Kocfinans
|
Features
|
The set of functions provided by JustICR Engine API can be divided into the following
groups:
1. Image input
|
Opening binary images from TIF files.
Reading binary images from memory.
Support of color and grayscale images from memory.
|
2. Training tool
|
Capture images from files or from the integrated application
Typing labels of each character
Verifying the suspected characters
Statistics tool
|
3. Image preprocessing
|
Deskewing
Despeckling
Lines removal
Rotation (90, 180, 270 degrees)
|
4. Recognition
|
Recognition is in field level
Handprint, machineprint, Farrington 7, CMC7
Special symbols and annotations can be trained and recognized
Support multilines fields
|
5. Advanced features
|
Dictionaries
Finding a text of specific font in a form
|
|
People who use JustICR say:
Training of a specific font
|
"We had a benchmark of a big application in the USA that
had an important OCR field of a specific font. This field was
so important, that the customer needed to have the highest recognition results with almost no
errors. The font was OCRA,
but it was of poor quality, because it was not the first copy of the form.
The best engines gave us 98% recognition with 1% false positives. After 1 day of training JustICR,
we achieved 99% recognition and no errors at all out of the 50,000 characters in
the benchmark." (2000, Airborne, USA)
|
Training of low quality writing
|
"The project went quite well until we installed the software, and our client faced the recognition quality of his forms. We
used three top ICR engines with advanced voting, but the results were so unsatisfactory (82% recognized characters and 5%
errors) that the client preferred not to use ICR at all. The writing quality was poor
because it was a second copy of the credit
card application; some of the writing was too light, and some was too bold. After one year, we decided to try JustICR.
We used voting of JustICR with one of the other engines that we had. The results jumped to 90% recognized characters and
2% errors." (1998-9, Unicre, Portugal)
|
Training of a new writing style and dictionary
|
The Turkish representative of a big software firm recounts his experience with JustICR:
"We decided to work with this software
because it was the only engine with the ability to learn the Turkish handwriting style and achieve
very good recognition quality."
The Brazilian Census, 1999: "The benchmark recognition requirements of this huge census
seemed to be impossible.
In order to be qualified, the system needed to achieve more than 90% recognition rate for
numeral ICR fields and more
than 30% recognition rate for alpha fields. The punishment for each field that was not
properly recognized was 30 times
bigger than the added points of recognized field.
During our tests, the recognition results for the numerals were above the minimal requirements using voting of 3 engines.
However, the situation with the alpha fields was a disaster. We had about 20% recognition rate per field, and 20% errors!
It looked impossible to achieve anything close to 30% errors per field with less than 1% errors.
And then we tried JustICR. We designed a form with several fields and gave it to 200 people. After collecting these forms
we trained JustICR in one day. We added JustICR into our voting system. We used a smart dictionary option that can
handle a partial dictionary with more than 25,000 words. The results were overwhelming: the alpha recognition rate was
57% and 0.3% errors per field, and the numeral field results were improved as well. Needless to say, we were the only
competitor in the benchmark who achieved the minimal requirements."
|
Looking for a font in a form
|
"The application that we needed to process had several hundreds of form variants.
Our form identification algorithm could
not deal with so many types of forms.
However, on each form there was a printed number in a specific font that indicated
the form type.
We used JustICR's ability to find and recognize this number, and the system worked flawlessly.
The competitors used various other approaches: one of them keyed in the information from image,
while another spent several hours, manually sorting the forms." (2000, IRS, Mexico)
The same technique was applied in several large applications in the USA,
Germany and India.
|
|
|
| |
|
|
