Do you ever struggle to read a friend's handwriting? Count yourself lucky,
then, that you're not working for the US Postal Service, which has to
decode and deliver something like 30 million handwritten envelopes
every single day!
With so much of our lives computerized, it's vitally important that machines and
humans can understand one another and pass information back and forth.
Mostly computers have things
their way—we have to "talk" to them through relatively crude devices such as keyboards
and mice so they can figure out what we want them to do. But when
it comes to processing more human kinds of information, like an
old-fashioned printed book or a letter scribbled with a
computers have to work much harder.
That's where optical character
recognition (OCR) comes in. It's a type of
software (program) that can automatically analyze printed text and turn it into
a form that a computer can process more easily. OCR is at the heart
of everything from handwriting analysis programs on cellphones to
the gigantic mail-sorting machines that ensure all those millions
of letters reach their destinations. How exactly does it work? Let's
take a closer look!
Photo: Recognizing characters#1: Can you make out the blue letter "P," beginning the word "Petrus," in this illuminated, hand-written bible dating from 1407CE? Imagine what a computerized optical character recognition program would make of it! Photo courtesy of
When it comes to optical character recognition, our brains and eyes are far superior to
As you read these words on your computer screen, your eyes and brain are
carrying out optical character recognition without you even noticing!
Your eyes are recognizing the patterns of light and dark that make up
the characters (letters, numbers, and things like punctuation
marks) printed on the screen and your brain is using those to figure
out what I'm trying to say (sometimes by reading individual
characters but mostly by scanning entire words and whole groups of
words at once).
Computers can do this too, but it's really hard work for them. The first
problem is that a computer has no eyes, so if you want it to read
something like the page of an old book, you have to present it with
an image of that page generated with an optical
scanner or a digital camera. The page you create this way is a
graphic file (often in the form of a JPG) and, as far as a computer's
concerned, there's no difference between it and a photograph of the
Taj Mahal or any other graphic: it's a completely meaningless pattern
of pixels (the colored dots or squares that make up any
computer graphic image). In other words, the computer has a picture of the page rather than the text itself—it can't read the words on the
page like we can, just like that. OCR is the process of turning a
picture of text into text itself—in other words, producing something
like a TXT or DOC file from a scanned JPG of a printed or handwritten
Photo: Recognizing characters: To you and me, it's the word "an", but to a computer this is just a meaningless pattern of black and white. And notice how the fibers in the paper are introducing some confusion into the image. If the ink were slightly more faded, the gray and white pattern of fibers would start to interfere and make the letters even harder to recognize.
What's the advantage of OCR?
Once a printed page is in this machine-readable text form, you can do
all kinds of things you couldn't do before. You can search through it
by keyword (handy if there's a huge amount of it), edit it with a
word processor, incorporate it into a Web page, compress it into a
ZIP file and store it in much less space, send it by email—and all
kinds of other neat things. Machine-readable text can also be decoded
by screen readers, tools that use speech synthesizers (computerized
voices, like the one Stephen Hawking used) to read out the words on a screen so
blind and visually impaired people can understand them. (Back in the
1970s, one of the first major uses of OCR was in a photocopier-like
device called the Kurzweil Reading Machine, which could read printed
books out loud to blind people.)
Photo: Scanning in your pocket: smartphone OCR apps are fast, accurate, and convenient. Left: Here I'm scanning the text of the article you're reading now, straight off my computer screen, with my smartphone and Text Scanner (an Android app by Peace). Right: A few seconds later, a very accurate version of the scanned text appears on my phone screen.
How does OCR work?
Let's suppose life was really simple and there was only one letter in the
alphabet: A. Even then, you can probably see that OCR would be quite
a tricky problem—because every single person writes the letter A in
a slightly different way. Even with printed text, there's an issue,
because books and other documents are printed in many different
typefaces (fonts) and the letter A can be printed in many subtly
Photo: There's a fair bit of variation between these different versions of a capital letter A,
printed in different computer fonts, but there's also a basic similarity: you can see that almost all of them are made from two angled lines that meet in the middle at the top, with a horizontal line between.
Broadly speaking, there are two different ways to
solve this problem, either by recognizing characters in their entirety
(pattern recognition) or by detecting the individual lines and
strokes characters are made from (feature detection) and identifying
them that way. Let's look at these in turn.
If everyone wrote the letter A exactly the same way, getting a computer
to recognize it would be easy. You'd just compare your scanned image
with a stored version of the letter A and, if the two matched, that
would be that. Kind of like Cinderella: "If the slipper fits..."
So how do you get everyone to write the same way? Back in the 1960s,
a special font called OCR-A was developed that could be used on
things like bank checks and so on. Every letter was exactly the same
width (so this was an example of what's called a monospace font) and
the strokes were carefully designed so each letter could easily be
distinguished from all the others. Check-printers were designed so
they all used that font, and OCR equipment was designed to recognize
it too. By standardizing on one simple font, OCR became a relatively
easy problem to solve. The only trouble is, most of what the world
prints isn't written in OCR-A—and no-one uses that font for their
handwriting! So the next step was to teach OCR programs to recognize
letters written in a number of very common fonts (ones like Times,
Helvetica, Courier, and so on). That meant they could recognize quite
a lot of printed text, but there was still no guarantee they could
recognize any font you might send their way.
Photo: OCR-A font: Designed to be read by computers as well as people. You might not recognize the style of text, but the numbers probably do look familiar to you from checks and computer printouts. Note that
similar-looking characters (like the lowercase "l" in Explain and the number "1" at the bottom) have been designed so computers can easily tell them apart.
Also known as feature extraction or intelligent character recognition
(ICR), this is a much more sophisticated way of spotting characters.
Suppose you're an OCR computer program presented with lots of
different letters written in lots of different fonts; how do you pick
out all the letter As if they all look slightly different? You could
use a rule like this: If you see two angled lines that meet in a
point at the top, in the center, and there's a horizontal line
between them about halfway down, that's a letter A. Apply that rule
and you'll recognize most capital letter As, no matter what font
they're written in. Instead of recognizing the complete pattern of an
A, you're detecting the individual component features (angled lines,
crossed lines, or whatever) from which the character is made. Most
modern omnifont OCR programs (ones that can recognize printed text
in any font) work by feature detection rather than pattern
recognition. Some use neural networks (computer programs
that automatically extract patterns in a brain-like way).
Photo: Feature detection: You can be pretty confident you're looking at a capital letter A if you can identify these three component parts joined together in the correct way.
How does handwriting recognition work?
Recognizing the characters that make up neatly laser-printed computer text is
relatively easy compared to decoding someone's scribbled
handwriting. That's the kind of simple-but-tricky, everyday problem
where human brains beat clever computers hands-down: we can all make
a rough stab at guessing the message hidden in even the worst human
writing. How? We use a combination of automatic pattern recognition,
feature extraction, and—absolutely crucially—knowledge about the writer and
the meaning of what's being written ("This letter, from my friend Harriet, is
about a classical concert we went to together, so the word she's
written here is more likely to be 'trombone' than 'tramline'.")
Cursive handwriting (with letters joined up and flowing together) is very much harder for a computer to recognize than computer-printed type, because it's difficult to know where one letter ends and another begins.
Many people write so hastily that they don't bother to form their letters fully, making recognition by pattern or feature extremely hard.
Another problem is that handwriting is an expression of individuality, so people may go out of their way to make their writing different from the norm.
When it comes to reading words like this, we rely heavily on the meaning of what's written, our knowledge of the writer, and the words that we've already read—something computers can't manage so easily.
Photo: The tricky problem of handwriting recognition.
Making it easy
Where computers do have to recognize handwriting, the problem is often
simplified for them. For example, mail-sorting computers generally
only have to recognize the zipcode (postcode) on an envelope, not the
entire address. So they just have to identify a relatively small
amount of text made only from basic letters and numbers. People are
encouraged to write the codes legibly (leaving spaces between
characters, using all uppercase letters) and, sometimes, envelopes
are preprinted with little squares for you to write the characters in to
help you keep them separate.
Forms designed to be processed by OCR sometimes have separate boxes for people to write each letter in or
faint guidelines known as comb fields, which encourage people to keep letters separate and
write legibly. (Generally the comb fields are printed in a special
color, such as pink, called a dropout color, which can be easily separated from the
text people actually write, usually in black or blue ink.)
Artwork: Forms designed for OCR incorporate simple aids to reduce scanning errors, including comb fields (top) and character boxes (middle) printed in a dropout color (pink),
and bubble choice fields or tick boxes (bottom).
Tablet computers and cellphones that have handwriting recognition often use feature extraction to recognize
letters as you write them. If you're writing a letter A, for example,
the touchscreen can sense you writing first one angled line, then the
other, and then the horizontal line joining them. In other words, the
computer gets a headstart in recognizing the features because you're
forming them separately, one after another, and that makes feature extraction much
easier than having to pick out the features from handwriting scribbled
What does OCR involve in practice?
Most people don't have to use OCR on an industrial scale: not many of us
will need to scan even a hundred documents a day, never mind a
million. We're more likely to want to OCR the occasional printed
article and turn it into editable form, or scan an out-of-copyright
book and republish it as a Web page.
Photo: What OCR looks like in practice: you load an image of a scanned page (the words in the background are a scanned bitmap), run your OCR program to generate a text file, and then step through the "recognized" text manually correcting errors. Here, in a neat twist of history, I'm using the
Linux Kooka scanning program to OCR the patent of Gustav Tauschek's original OCR system from 1928/1929 (described below).
In practice, this is what everyday OCR actually involves:
Printout: First, you need to get the best possible printout of your existing document.
Sometimes you're just stuck with an old typewritten script, but you
may be able to improve the print quality by photocopying (to
increase the contrast between the print and the page). The quality
of the original printout makes a huge difference to the accuracy of
the OCR process. Dirty marks, folds, coffee stains, ink blots, and
any other stray marks will all reduce the likelihood of correct
letter and word recognition.
Scanning: You run the printout through your optical scanner. Sheet-feed scanners
are better for OCR than flatbed scanners because you can scan pages one
after another. Most modern OCR programs will scan each page,
recognize the text on it, and then scan the next page automatically.
If you're using a flatbed scanner, you'll have to insert the pages
one at a time by hand. If you've got a reasonably good digital
camera, you may be able to create images of your pages by taking
photos. You'll probably need to use a macro (close-up) focus setting to get
really sharp letters that are clear enough for accurate OCR.
Two-color: The first stage in OCR involves generating a black-and-white
(two-color/one-bit) version of the color or grayscale scanned page,
similar to what you'd see coming out of a fax machine. OCR is
essentially a binary process: it recognizes things that are either
there or not. If the original scanned image is perfect, any black it
contains will be part of a character that needs to be recognized
while any white will be part of the background. Reducing the image
to black and white is therefore the first stage in figuring out the
text that needs processing—although it can also introduce errors. If you
have a color scan of a newspaper with a large brown coffee stain
over the words, it's easy to tell the text from the stain; but if
you reduce the scan to a black-and-white image, the stain will turn
to black and white too and may confuse the OCR process.
OCR: All OCR programs are slightly different, but generally they
process the image of each page by recognizing the text character by character,
word by word, and line by line. In the mid-1990s, OCR programs were so slow that
you could literally watch them "reading" through and processing the
text while you waited; computers are far faster now and OCR is
pretty much instantaneous.
Basic error correction: Some programs give you the opportunity to review and correct each page in
turn: they instantly process the entire page and then use a built-in
spellchecker to highlight any apparently misspelled words that may
indicate a misrecognition, so you can automatically correct the
mistake. You can usually switch off this feature if you want to, if
you have many pages to scan and you don't want to check them all as
you're going along. Sophisticated OCR programs have extra error
checking features to help you spot mistakes. For example, some use
what's called near-neighbor analysis to find words that are likely
to occur nearby, so text incorrectly recognized as "the
barking bog" might be automatically changed to "the barking dog"
(because "barking" and "dog" are two words that very often
Layout analysis: Good OCR programs automatically detect complex page layouts, such as
multiple columns of text, tables, images, and so on. Images are automatically turned
into graphics, tables are (with luck) turned into tables, and
columns are split up correctly (so the text from the first line of
the first column isn't automatically joined to the text from the
first line of the second column).
Proofreading: Even the best OCR programs aren't perfect, especially when they're
working from very old documents or poor quality printed text. That's
why the final stage in OCR should always be a good, old-fashioned
This is how Google's OCR process mangled it—and how Google still publishes it today, with no
attempt at human editing or correction. Not only are the errors uncorrected, but text from different
points of the document has been added out of order, making the whole thing meaningless and impossible to understand.
Always, always, always check and correct the output from OCR before publishing it.
Who invented OCR?
Most people think getting machines to read human text is a relatively
recent innovation, but it's older than you might suppose. Here's a
whistle-stop tour through OCR history:
1914: Edward Fournier D'Albe describes the Optophone, a machine for reading documents aloud to blind people, by scanning printed words and converting them into musical tones.
1928/9: Gustav Tauschek of Vienna, Austria
patents a basic OCR "reading machine." Paul Handel of General Electric files a patent for a similar system in the United States in April 1931.
Both are based on the idea of using light-detecting photocells to recognize patterns on paper or card.
1949: L.E. Flory and W.S. Pike of RCA Laboratories develop a photocell-based machine that can read text to blind people at a
rate of 60 words per minute. (Read all about it in the February 1949 issue of Popular Science.)
1950: David H. Shepard develops machines that can turn printed information into machine-readable form for the US military and later
founds a pioneering OCR company called Intelligent Machines
Research (IMR). Shepherd also develops a machine-readable font called Farrington B (also called OCR-7B and 7B-OCR), now widely used to print the embossed numbers on credit cards.
1960: Lawrence (Larry) Roberts, a computer graphics researcher working at MIT, develops early text recognition using specially simplified fonts such as OCR-A. He later becomes one of the
founding fathers of the Internet.
1950s/1960s: Reader's Digest and RCA work together to develop some of the first commercial OCR systems.
1960s: Postal services around the world begin to use OCR technology for mail-sorting. They include the US Postal Service, Britain's General Post Office (GPO, now called Royal Mail), Canada Post, and the
German Deutsche Post. Helped by companies such as Lockheed Martin, postal services remain at the forefront of OCR research to this day.
1974: Raymond Kurzweil develops the Kurzweil Reading Machine (KRM) that combines a flatbed scanner and speech synthesizer in a machine that
can read printed pages aloud to blind people. Kurzweil's OCR software is acquired by Xerox and marketed under the names ScanSoft and (later) Nuance Communications.
1980s: Hewlett-Packard begins work on Tesseract, now an open-source OCR project sponsored by Google
that can work with text in 60 languages.
1993: The Apple Newton MessagePad (PDA) is one of the first handheld computers to feature handwriting recognition on a touch-sensitive screen. During the 1990s,
handwriting recognition becomes an increasingly popular feature on cellphones, PDAs (notably the pioneering
PalmPilot), and other handhelds.
2000: Researchers at Carnegie Mellon University flip the problem of developing a good OCR
system on its head—and develop a spam-busting system called CAPTCHA (see caption below).
Photo: We know from OCR research that computers find it hard to recognize badly printed words that humans can read relatively easily. That's why CAPTCHA puzzles like this are used to stop spammers from bombarding email systems, message boards, and other websites. This one was developed by Carnegie Mellon University and later acquired by Google as part of its original reCAPTCHA system. The original reCAPTCHA had an added benefit: when you type in the garbled words, you helped Google to recognize part of the scanned text from an old book that it wanted to convert to machine-readable form. In effect, you were doing a little bit of OCR on Google's behalf. Most websites have now switched to a different, more secure CAPTCHA test that involves identifying photos of cars, mountains, fire hydrants, and other everyday things.
2007: The arrival of the iPhone prompts the development of handy, point-and-click smartphone apps that can scan and convert text using a phone camera.
2019: Amazon announces the release of
Textract, a cloud-based OCR system heavily based on
machine learning, which can cope with
forms, spreadsheets, invoices written in different styles, and all kinds of other business documents.
Find out more
On this website
You might like these other articles on our site covering related topics:
A picture of a thousand words? by Evin Levey. Google Blog, October 30, 2008. How Google converted scanned PDFs into indexable text to make its search results more useful.
If you like technical details, you'll find OCR patents worth a look. Here a few representative examples to start you off; you'll find many more on Google Patents (you can use the search operator "intitle:OCR" to find hundreds of relevant patents).
Reading system by Raymond Kurzweil, published December 23, 1999. One of Ray Kurzweil's later OCR patents describing a computerized reading machine for the blind that scans a printed page, turns it into a text file, and then reads the text out loud.
Please do NOT copy our articles onto blogs and other websites
Articles from this website are registered at the US Copyright Office. Copying or otherwise using registered works without permission, removing this or other copyright notices, and/or infringing related rights could make you liable to severe civil or criminal penalties.