Slashdot Mirror
Where Did Affordable OCR Go?
Goeland86 asks: "Has OCR (Optical Character Recognition) died down? Where have all the magical programs that translate your handwriting to office compatible files gone? Most of the windows programs nowadays are either expensive (ReadIris Pro 9 about $400) and not that many OSS projects for OCR have released a recent update (Kognition was last updated on July 17th 2003 according to Freshmeat). Has everyone already scanned/translated all of their paper files? Has OCR outlived its use, or is it just a fancy technology that hit a dead end in terms of the market? Have Slashdot readers used it? If so, are you still using it? If not, why?"
I usually don't post replies like this, but this question is ridiculously underresearched. OCR is a hard problem. Sure, a OSS alternative would be nice, but until a solution matures, when you really need OCR you need it because it's generally unreasonable either from a time standpoint or a budget standpoint to any alternative. That is why people pay for software sometimes.
TextBridge, PaperPort, and a host of other entry level programs are available for windows under a $100 price point. Generally if you buy a decent scanner (ie not a $50 piece of crap), you'll get some software capable of doing OCR bundled for free.
Higher-end OCR packages with better accuracy, more features, etc. often cost quite a bit more. OmniPage Pro is a decent package for only slightly more than $100. ReadIris is a really good program, and is reportedly very quick in comparison to some of the others. I imagine this is the reason that it costs $400.
There are document management packages out there that have very good OCR integrated that cost a hell of a lot more than $400. Trust me, though, if you're looking at the time or cost of converting a few thousand pages of data into editible text documents, a program that costs even $400 should be a steal.
My company just paid ~$1,800USD for OCR software (ABBYY FormReader). We're scanning in stacks of healthcare forms, reading the data, and spitting them out into DBF format. Why? I don't know, I just do my job. It was my responsiblity to review and demo other pieces of software, and ABBYY's was definitely the most robust. Open Source had, as the poster stated, few contenders and even fewer that had been worked on since the 90s.
I don't know what happened to OCR, but there's certainly still a need for it.
El riesgo vive siempre!
It does a "Feeling Lucky" search since http by itself is not a valid url. Go search on Google for http and look at the first result.
The DP site does OCR and proofreads the results for Project Gutenberg. Anyone can join and spend a few minutes once in a while proofreading books. If you are kind of ADD like me, it lets you read about 3 pages of a book once in a while without having to actually sit down and do cover-to-cover.
When you scan to a PDF, you essentially create a high-resolution single channel JPEG image which is the sole contents of the PDF file.
It does not create text that you can search or highlight and copy with your mouse later on. It's just a picture.
Now, there is some nice scanning software out there that if you do select "text" mode when you scan to PDF, it does an OCR pass and sticks that in the PDF. But the cost of this software is usually hidden in the purchase of a high-end scanner or printer/fax combo that it would normally accompany.
THIS THING CAN TURN ON A DIME, MACROSSZERO STYLE ALSO FUCK BETA, ~NYORON
(I'm in the document imaging / conversion industry)
:)
The term paperless office is considered a joke, and the funny part of it is this: as soon as someone looks up a document in their doc management system they just print it. Even if just to glance at! Copier/printer companies are thrilled!
There are megatons of paper and microfilm out there left to ocr and process. It's considered a pretty fast growing industry, although stunted recently after the bomb and more by the economy.
Having ocr'd images is very handy. Here's an open secret though-- Image+ hidden text pdf.
--Searchable, you have the original doc just as it looked, and the ocr errors don't make such an impact. It's easy to throw into a search engine and the prints look great, and small (b+w use tiff group iv, and jpeg for color jbig is not quite mature yet and only a few apps from cvision do a great job at it)
Anyway, since people just hit print as soon as they find their doc in a system those file cabinets we tried so hard to empty and organize re fill magically.
Also, scanning and setting up an edms (electonic doc managemnt system) is considered a luxury. business move slow with luxury items and usually get to reap the benefits of more mature software and systems (but this is NOT always true!).
Many other slow tech adoption business are just discovering scanning ocr and doc management. Litigation is a great example. xerox was doing quite a few tv ads recently touting that stuff.
The state of ocr itself is strange. There has been a sort of pleague in that industry of 'weird innovation' for years and many buyouts or companys changing the focus of their ocr product to another industry (like web or xml). Even the small office versions ($500 range) are not geared for any sort of reasonable volume or speed without crashing and burning, and usually designed to be babysat. Using these apps leaves the user with a really bad experience. For those not familliar the process goes something like this for a 200 page b+w document:
==
Scan (or import but import is usually crippled)
gaze at loads of memory hogging eye candy (this is what your upgrade bought you usually)
wait
correct skew (wait for crappy tools)
(possibly reboot from crash)
recognize page -slower with each new version even
when hardware is so much faster every year. some recognition is improved in some packages. Some of the latest i've tested take over 15 sec and sometimes over 45 sec per page!
(crash!?)
Correct errors / tune learning engine. (sometimes i swear this effort of teaching goes straight to NULL)
repeat 199 times
Now since you're locked in your desk and finished scanning now it's time to export! (like i didn't know what formats i wanted before i sat down.)
So it chews and chews and maybe crashes causing you to repeat all the above steps. Also note that most of these apps keep all the pages pretty much uncompressed in memory, then create a copy of them in memory for your desired output format. (crash)
2 days of work gone.
====
Most users walk away with the feeling of 'Yikes! all I wanted was a word doc of this. I'll just do something else'
For the home and also small biz market here are some of the 'weird innovations'--
===========
typereader 5 -- pretty good app! doesn't do image+hidden text pdf though. Pitty. has a batch file import and reasonable priced in the $100 range. nice and fast with good results
Typereader 6 and up- file import feature moved to industrial version lots of eye candy less stable minor improvement in recog and a bunch of other silly limits & slow
Omnipage same thing only it's never been great for over 50 pages. horrid workflow and crashes like crazy. very unpredictable!
Omnipage version 3 was better in many ways than omnipage 14. (lightning fast on today's equipment too
abby finereader - very slow but great recognition, more stable but lame workflow-
Firefox &
It's slightly off topic but seemed appropriate.
.. // c d
Here's some quick tips/nuggets of crispy wisdom.
The art of ocr is like working with autistics. give them what they expect. the more surprises, the more episodes.
Don't believe the hype.
Scan black & white to TIFF GROUP IV. OCR systems are optomized for this. Color is new and pretty wacky still. BMP even freaks out in black and white on some packages.
Make sure your background is white and clean, not specled. despeckling tools can be overused and kill ocr results.
3 hole punches regularly show up as o O 0 D
staples: ~
Deskew all images to a line of text, not the page
Scan at 200-300 dpi but not higher than 600 or most apps will choke and produce bad results.
Make a custom dictionary if you can. if you're doing automotive related stuff, look up auto terms and make a dictionary out of it.
To process tiny text (concordences etc) scan at 800dpi and then fool the ocr by scaling the image to 300. sounds nuts right? ok try it the logcal way first and then come back and try this teq.
Shaded text is a new thing in document as is inverted text blocks (thanks word...you make my job hell.)you must remove the shading with something like scanfix by tms sequioa- good tool for small doc cleanup for pre-ocr. requires practice and trial and error. interface needs some work though.
Dot matrix prints should be scanned and some blur added to join the dots (unless you are using something expressedly made for dmp's) as always your milage may vary
Turn off auto rotate (mangle)features. They are not very smart an often have monkeyvision. just review your images before hand and rotate accordingly.
If you're scanning something poster size, or engineering drawing size (not recommended for most ocr) cut it into smaller images. ideally regions of interest not larger than 8.5x11
Remember 99% accurate means 1 character per hundred will be screwed up.
Table of contents pages are an interesting test for ocr especially if they use periods to lead to page numbers. How many identical characters can occur before the ocr system misreads. often quite telling.
OCRing a spreadsheet and using the data with out verifying every character? may the monkeygods help you.
Above applies to processing screenshots, 17'th century print, tabloid print, multi-column, shaded background handwriting w/o special software, modern magazines, etc.
OCR does not like non seriffed fonts much.
The post office spends millions on theirs and they have a nice address DB list to verify against.
Over 90% of banks scan your checks and microfilm them. (and there is some really cool signature verification sofware out there for forgery detection) MICR font at the bottom helps them immensely.
Breaking down your document into areas can be useful. changing fonts and sizes sometimes throw it off . an example would be computer lit with code snippets interspersed.
Do yourself a favor if it applies and use image+hidden text pdf. raw ocr is almost always yucky and all those claims of preserving document layout and format are just that--claims.
If you do use i+HT pdf, or for a larger job for that matter, do it in small chunks so your app doesn't crash. for pdf, join the small documents together in acrobat later or use other tools to do so.
For fun and science, take an old apple newton 100 and trace over some of the text on your page and compare its results to your ocr package.
Anyway i hope that helps someone avoid a few landmines and there are many more tips out there. these are from my experience and off the cuff.
Firefox &