<s>
A	O
process	O
called	O
Optical	B-Application
Character	I-Application
Recognition	I-Application
(	O
OCR	O
)	O
converts	O
printed	O
texts	O
into	O
digital	O
image	O
files	O
.	O
</s>
<s>
It	O
is	O
a	O
digital	O
copier	O
that	O
uses	O
automation	O
to	O
convert	O
scanned	O
documents	O
into	O
editable	O
,	O
shareable	O
PDFs	B-Application
that	O
are	O
machine-readable	O
.	O
</s>
<s>
OCR	B-Application
software	I-Application
can	O
extract	O
data	O
from	O
scanned	O
documents	O
,	O
camera	O
photos	O
,	O
and	O
image-only	O
PDFs	B-Application
.	O
</s>
<s>
It	O
makes	O
static	O
material	O
editable	O
and	O
does	O
away	O
with	O
the	O
necessity	O
for	O
human	O
data	B-General_Concept
entry	I-General_Concept
.	O
</s>
<s>
Widely	O
used	O
as	O
a	O
form	O
of	O
data	B-General_Concept
entry	I-General_Concept
from	O
printed	O
paper	O
data	O
records	O
–	O
whether	O
passport	O
documents	O
,	O
invoices	O
,	O
bank	O
statements	O
,	O
computerized	O
receipts	O
,	O
business	O
cards	O
,	O
mail	O
,	O
printouts	O
of	O
static-data	O
,	O
or	O
any	O
suitable	O
documentation	O
–	O
it	O
is	O
a	O
common	O
method	O
of	O
digitizing	O
printed	O
texts	O
so	O
that	O
they	O
can	O
be	O
electronically	O
edited	O
,	O
searched	O
,	O
stored	O
more	O
compactly	O
,	O
displayed	O
on-line	O
,	O
and	O
used	O
in	O
machine	O
processes	O
such	O
as	O
cognitive	B-General_Concept
computing	I-General_Concept
,	O
machine	B-Application
translation	I-Application
,	O
(	O
extracted	O
)	O
text-to-speech	B-Application
,	O
key	O
data	B-Algorithm
and	I-Algorithm
text	I-Algorithm
mining	I-Algorithm
.	O
</s>
<s>
OCR	O
is	O
a	O
field	O
of	O
research	O
in	O
pattern	O
recognition	O
,	O
artificial	B-Application
intelligence	I-Application
and	O
computer	B-Application
vision	I-Application
.	O
</s>
<s>
Early	O
optical	B-Application
character	I-Application
recognition	I-Application
may	O
be	O
traced	O
to	O
technologies	O
involving	O
telegraphy	O
and	O
creating	O
reading	O
devices	O
for	O
the	O
blind	O
.	O
</s>
<s>
This	O
device	O
required	O
the	O
invention	O
of	O
two	O
enabling	O
technologiesthe	O
CCD	B-Algorithm
flatbed	O
scanner	O
and	O
the	O
text-to-speech	B-Application
synthesizer	I-Application
.	O
</s>
<s>
In	O
1978	O
,	O
Kurzweil	O
Computer	O
Products	O
began	O
selling	O
a	O
commercial	O
version	O
of	O
the	O
optical	B-Application
character	I-Application
recognition	I-Application
computer	O
program	O
.	O
</s>
<s>
LexisNexis	B-Library
was	O
one	O
of	O
the	O
first	O
customers	O
,	O
and	O
bought	O
the	O
program	O
to	O
upload	O
legal	O
paper	O
and	O
news	O
documents	O
onto	O
its	O
nascent	O
online	O
databases	O
.	O
</s>
<s>
In	O
the	O
2000s	O
,	O
OCR	O
was	O
made	O
available	O
online	O
as	O
a	O
service	O
(	O
WebOCR	O
)	O
,	O
in	O
a	O
cloud	B-Architecture
computing	I-Architecture
environment	O
,	O
and	O
in	O
mobile	O
applications	O
like	O
real-time	O
translation	O
of	O
foreign-language	O
signs	O
on	O
a	O
smartphone	B-Application
.	O
</s>
<s>
With	O
the	O
advent	O
of	O
smart-phones	B-Application
and	O
smartglasses	B-General_Concept
,	O
OCR	O
can	O
be	O
used	O
in	O
internet	O
connected	O
mobile	O
device	O
applications	O
that	O
extract	O
text	O
captured	O
using	O
the	O
device	O
's	O
camera	O
.	O
</s>
<s>
These	O
devices	O
that	O
do	O
not	O
have	O
OCR	O
functionality	O
built	O
into	O
the	O
operating	O
system	O
will	O
typically	O
use	O
an	O
OCR	O
API	B-Application
to	O
extract	O
the	O
text	O
from	O
the	O
image	O
file	O
captured	O
and	O
provided	O
by	O
the	O
device	O
.	O
</s>
<s>
The	O
OCR	O
API	B-Application
returns	O
the	O
extracted	O
text	O
,	O
along	O
with	O
information	O
about	O
the	O
location	O
of	O
the	O
detected	O
text	O
in	O
the	O
original	O
image	O
back	O
to	O
the	O
device	O
app	O
for	O
further	O
processing	O
(	O
such	O
as	O
text-to-speech	B-Application
)	O
or	O
display	O
.	O
</s>
<s>
Various	B-Application
commercial	I-Application
and	I-Application
open	I-Application
source	I-Application
OCR	I-Application
systems	I-Application
are	O
available	O
for	O
most	O
common	O
writing	O
systems	O
,	O
including	O
Latin	O
,	O
Cyrillic	O
,	O
Arabic	O
,	O
Hebrew	O
,	O
Indic	O
,	O
Bengali	O
(	O
Bangla	O
)	O
,	O
Devanagari	O
,	O
Tamil	O
,	O
Chinese	O
,	O
Japanese	O
,	O
and	O
Korean	O
characters	O
.	O
</s>
<s>
Data	B-General_Concept
entry	I-General_Concept
for	O
business	O
documents	O
,	O
e.g.	O
</s>
<s>
Optical	B-Application
character	I-Application
recognition	I-Application
(	O
OCR	O
)	O
targets	O
typewritten	O
text	O
,	O
one	O
glyph	B-Application
or	O
character	O
at	O
a	O
time	O
.	O
</s>
<s>
Intelligent	B-Application
character	I-Application
recognition	I-Application
(	O
ICR	O
)	O
also	O
targets	O
handwritten	O
printscript	O
or	O
cursive	O
text	O
one	O
glyph	B-Application
or	O
character	O
at	O
a	O
time	O
,	O
usually	O
involving	O
machine	O
learning	O
.	O
</s>
<s>
Intelligent	B-General_Concept
word	I-General_Concept
recognition	I-General_Concept
(	O
IWR	O
)	O
also	O
targets	O
handwritten	O
printscript	O
or	O
cursive	O
text	O
,	O
one	O
word	O
at	O
a	O
time	O
.	O
</s>
<s>
This	O
is	O
especially	O
useful	O
for	O
languages	O
where	O
glyphs	B-Application
are	O
not	O
separated	O
in	O
cursive	O
script	O
.	O
</s>
<s>
There	O
are	O
cloud	O
based	O
services	O
which	O
provide	O
an	O
online	O
OCR	O
API	B-Application
service	O
.	O
</s>
<s>
Handwriting	O
movement	O
analysis	O
can	O
be	O
used	O
as	O
input	O
to	O
handwriting	B-Application
recognition	I-Application
.	O
</s>
<s>
Instead	O
of	O
merely	O
using	O
the	O
shapes	O
of	O
glyphs	B-Application
and	O
words	O
,	O
this	O
technique	O
is	O
able	O
to	O
capture	O
motions	O
,	O
such	O
as	O
the	O
order	O
in	O
which	O
segments	O
are	O
drawn	O
,	O
the	O
direction	O
,	O
and	O
the	O
pattern	O
of	O
putting	O
the	O
pen	O
down	O
and	O
lifting	O
it	O
.	O
</s>
<s>
This	O
technology	O
is	O
also	O
known	O
as	O
"	O
on-line	O
character	B-Application
recognition	I-Application
"	O
,	O
"	O
dynamic	O
character	B-Application
recognition	I-Application
"	O
,	O
"	O
real-time	O
character	B-Application
recognition	I-Application
"	O
,	O
and	O
"	O
intelligent	B-Application
character	I-Application
recognition	I-Application
"	O
.	O
</s>
<s>
OCR	B-Application
software	I-Application
often	O
"	O
pre-processes	O
"	O
images	O
to	O
improve	O
the	O
chances	O
of	O
successful	O
recognition	O
.	O
</s>
<s>
BinarisationConvert	O
an	O
image	O
from	O
color	O
or	O
greyscale	O
to	O
black-and-white	O
(	O
called	O
a	O
"	O
binary	B-Algorithm
image	I-Algorithm
"	O
because	O
there	O
are	O
two	O
colors	O
)	O
.	O
</s>
<s>
The	O
task	O
of	O
binarisation	O
itself	O
is	O
necessary	O
since	O
most	O
commercial	O
recognition	O
algorithms	O
work	O
only	O
on	O
binary	B-Algorithm
images	I-Algorithm
since	O
it	O
proves	O
to	O
be	O
simpler	O
to	O
do	O
so	O
.	O
</s>
<s>
In	O
addition	O
,	O
the	O
effectiveness	O
of	O
the	O
binarisation	O
step	O
influences	O
to	O
a	O
significant	O
extent	O
the	O
quality	O
of	O
the	O
character	B-Application
recognition	I-Application
stage	O
and	O
the	O
careful	O
decisions	O
are	O
made	O
in	O
the	O
choice	O
of	O
the	O
binarisation	O
employed	O
for	O
a	O
given	O
input	O
image	O
type	O
;	O
since	O
the	O
quality	O
of	O
the	O
binarisation	O
method	O
employed	O
to	O
obtain	O
the	O
binary	O
result	O
depends	O
on	O
the	O
type	O
of	O
the	O
input	O
image	O
(	O
scanned	O
document	O
,	O
scene	B-Algorithm
text	I-Algorithm
image	O
,	O
historical	O
degraded	O
document	O
etc	O
.	O
</s>
<s>
Layout	B-Algorithm
analysis	I-Algorithm
or	O
"	O
zoning	O
"	O
Identifies	O
columns	O
,	O
paragraphs	O
,	O
captions	O
,	O
etc	O
.	O
</s>
<s>
Especially	O
important	O
in	O
multi-column	O
layouts	O
and	O
tables	B-Application
.	O
</s>
<s>
Matrix	O
matching	O
involves	O
comparing	O
an	O
image	O
to	O
a	O
stored	O
glyph	B-Application
on	O
a	O
pixel-by-pixel	O
basis	O
;	O
it	O
is	O
also	O
known	O
as	O
"	O
pattern	O
matching	O
"	O
,	O
"	O
pattern	O
recognition	O
"	O
,	O
or	O
"	O
image	O
correlation	O
"	O
.	O
</s>
<s>
This	O
relies	O
on	O
the	O
input	O
glyph	B-Application
being	O
correctly	O
isolated	O
from	O
the	O
rest	O
of	O
the	O
image	O
,	O
and	O
on	O
the	O
stored	O
glyph	B-Application
being	O
in	O
a	O
similar	O
font	O
and	O
at	O
the	O
same	O
scale	O
.	O
</s>
<s>
Feature	O
extraction	O
decomposes	O
glyphs	B-Application
into	O
"	O
features	O
"	O
like	O
lines	O
,	O
closed	O
loops	O
,	O
line	O
direction	O
,	O
and	O
line	O
intersections	O
.	O
</s>
<s>
These	O
features	O
are	O
compared	O
with	O
an	O
abstract	O
vector-like	O
representation	O
of	O
a	O
character	O
,	O
which	O
might	O
reduce	O
to	O
one	O
or	O
more	O
glyph	B-Application
prototypes	O
.	O
</s>
<s>
General	O
techniques	O
of	O
feature	O
detection	O
in	O
computer	B-Application
vision	I-Application
are	O
applicable	O
to	O
this	O
type	O
of	O
OCR	O
,	O
which	O
is	O
commonly	O
seen	O
in	O
"	O
intelligent	O
"	O
handwriting	B-Application
recognition	I-Application
and	O
indeed	O
most	O
modern	O
OCR	B-Application
software	I-Application
.	O
</s>
<s>
Nearest	B-General_Concept
neighbour	I-General_Concept
classifiers	I-General_Concept
such	O
as	O
the	O
k-nearest	B-General_Concept
neighbors	I-General_Concept
algorithm	I-General_Concept
are	O
used	O
to	O
compare	O
image	O
features	O
with	O
stored	O
glyph	B-Application
features	O
and	O
choose	O
the	O
nearest	O
match	O
.	O
</s>
<s>
Software	O
such	O
as	O
Cuneiform	B-Language
and	O
Tesseract	B-Language
use	O
a	O
two-pass	O
approach	O
to	O
character	B-Application
recognition	I-Application
.	O
</s>
<s>
Modern	O
OCR	B-Application
software	I-Application
include	O
Google	B-Application
Docs	I-Application
OCR	O
,	O
ABBYY	B-Operating_System
FineReader	I-Operating_System
and	O
Transym	O
.	O
</s>
<s>
Others	O
like	O
OCRopus	B-Language
and	O
Tesseract	B-Language
uses	O
neural	B-Architecture
networks	I-Architecture
which	O
are	O
trained	O
to	O
recognize	O
whole	O
lines	O
of	O
text	O
instead	O
of	O
focusing	O
on	O
single	O
characters	O
.	O
</s>
<s>
The	O
OCR	O
result	O
can	O
be	O
stored	O
in	O
the	O
standardized	O
ALTO	B-Language
format	O
,	O
a	O
dedicated	O
XML	O
schema	O
maintained	O
by	O
the	O
United	O
States	O
Library	O
of	O
Congress	O
.	O
</s>
<s>
Other	O
common	O
formats	O
include	O
hOCR	B-General_Concept
and	O
PAGE	O
XML	O
.	O
</s>
<s>
For	O
a	O
list	B-Application
of	I-Application
optical	I-Application
character	I-Application
recognition	I-Application
software	I-Application
see	O
Comparison	B-Application
of	I-Application
optical	I-Application
character	I-Application
recognition	I-Application
software	I-Application
.	O
</s>
<s>
This	O
might	O
be	O
,	O
for	O
example	O
,	O
all	O
the	O
words	O
in	O
the	O
English	O
language	O
,	O
or	O
a	O
more	O
technical	O
lexicon	B-Application
for	O
a	O
specific	O
field	O
.	O
</s>
<s>
This	O
technique	O
can	O
be	O
problematic	O
if	O
the	O
document	O
contains	O
words	O
not	O
in	O
the	O
lexicon	B-Application
,	O
like	O
proper	O
nouns	O
.	O
</s>
<s>
Tesseract	B-Language
uses	O
its	O
dictionary	O
to	O
influence	O
the	O
character	O
segmentation	O
step	O
,	O
for	O
improved	O
accuracy	O
.	O
</s>
<s>
The	O
output	O
stream	O
may	O
be	O
a	O
plain	O
text	O
stream	O
or	O
file	O
of	O
characters	O
,	O
but	O
more	O
sophisticated	O
OCR	O
systems	O
can	O
preserve	O
the	O
original	O
layout	O
of	O
the	O
page	O
and	O
produce	O
,	O
for	O
example	O
,	O
an	O
annotated	O
PDF	B-Application
that	O
includes	O
both	O
the	O
original	O
image	O
of	O
the	O
page	O
and	O
a	O
searchable	O
textual	O
representation	O
.	O
</s>
<s>
The	O
Levenshtein	O
Distance	O
algorithm	O
has	O
also	O
been	O
used	O
in	O
OCR	O
post-processing	O
to	O
further	O
optimize	O
results	O
from	O
an	O
OCR	O
API	B-Application
.	O
</s>
<s>
Beyond	O
an	O
application-specific	O
lexicon	B-Application
,	O
better	O
performance	O
may	O
be	O
had	O
by	O
taking	O
into	O
account	O
business	O
rules	O
,	O
standard	O
expression	O
,	O
or	O
rich	O
information	O
contained	O
in	O
color	O
images	O
.	O
</s>
<s>
This	O
strategy	O
is	O
called	O
"	O
Application-Oriented	O
OCR	O
"	O
or	O
"	O
Customized	O
OCR	O
"	O
,	O
and	O
has	O
been	O
applied	O
to	O
OCR	O
of	O
license	O
plates	O
,	O
invoices	O
,	O
screenshots	B-Application
,	O
ID	O
cards	O
,	O
driver	O
licenses	O
,	O
and	O
automobile	O
manufacturing	O
.	O
</s>
<s>
There	O
are	O
several	O
techniques	O
for	O
solving	O
the	O
problem	O
of	O
character	B-Application
recognition	I-Application
by	O
means	O
other	O
than	O
improved	O
OCR	O
algorithms	O
.	O
</s>
<s>
As	O
Google	B-Language
Tesseract	I-Language
can	O
be	O
trained	O
to	O
recognize	O
new	O
fonts	O
,	O
it	O
can	O
recognize	O
OCR-A	O
,	O
OCR-B	O
and	O
MICR	O
fonts	O
.	O
</s>
<s>
"	O
Comb	O
fields	O
"	O
are	O
pre-printed	O
boxes	O
that	O
encourage	O
humans	O
to	O
write	O
more	O
legiblyone	O
glyph	B-Application
per	O
box	O
.	O
</s>
<s>
These	O
are	O
often	O
printed	O
in	O
a	O
"	B-General_Concept
dropout	I-General_Concept
color	I-General_Concept
"	I-General_Concept
which	O
can	O
be	O
easily	O
removed	O
by	O
the	O
OCR	O
system	O
.	O
</s>
<s>
Palm	B-Operating_System
OS	I-Operating_System
used	O
a	O
special	O
set	O
of	O
glyphs	B-Application
,	O
known	O
as	O
"	O
Graffiti	B-Application
"	O
which	O
are	O
similar	O
to	O
printed	O
English	O
characters	O
but	O
simplified	O
or	O
modified	O
for	O
easier	O
recognition	O
on	O
the	O
platform	O
's	O
computationally	O
limited	O
hardware	O
.	O
</s>
<s>
Users	O
would	O
need	O
to	O
learn	O
how	O
to	O
write	O
these	O
special	O
glyphs	B-Application
.	O
</s>
<s>
Crowdsourcing	O
humans	O
to	O
perform	O
the	O
character	B-Application
recognition	I-Application
can	O
quickly	O
process	O
images	O
like	O
computer-driven	O
OCR	O
,	O
but	O
with	O
higher	O
accuracy	O
for	O
recognizing	O
images	O
than	O
that	O
obtained	O
via	O
computers	O
.	O
</s>
<s>
The	O
National	O
Library	O
of	O
Finland	O
has	O
developed	O
an	O
online	O
interface	O
for	O
users	O
to	O
correct	O
OCRed	O
texts	O
in	O
the	O
standardized	O
ALTO	B-Language
format	O
.	O
</s>
<s>
Crowd	O
sourcing	O
has	O
also	O
been	O
used	O
not	O
to	O
perform	O
character	B-Application
recognition	I-Application
directly	O
but	O
to	O
invite	O
software	O
developers	O
to	O
develop	O
image	O
processing	O
algorithms	O
,	O
for	O
example	O
,	O
through	O
the	O
use	O
of	O
rank-order	O
tournaments	O
.	O
</s>
<s>
One	O
study	O
based	O
on	O
recognition	O
of	O
19th	O
-	O
and	O
early	O
20th-century	O
newspaper	O
pages	O
concluded	O
that	O
character-by-character	O
OCR	O
accuracy	O
for	O
commercial	O
OCR	B-Application
software	I-Application
varied	O
from	O
81%	O
to	O
99%	O
;	O
total	O
accuracy	O
can	O
be	O
achieved	O
by	O
human	O
review	O
or	O
Data	O
Dictionary	O
Authentication	O
.	O
</s>
<s>
The	O
MNIST	B-General_Concept
database	I-General_Concept
is	O
commonly	O
used	O
for	O
testing	O
systems	O
 '	O
ability	O
to	O
recognise	O
handwritten	O
digits	O
.	O
</s>
<s>
For	O
example	O
,	O
if	O
word	O
context	O
(	O
basically	O
a	O
lexicon	B-Application
of	O
words	O
)	O
is	O
not	O
used	O
to	O
correct	O
software	O
finding	O
non-existent	O
words	O
,	O
a	O
character	O
error	O
rate	O
of	O
1%	O
(	O
99%	O
accuracy	O
)	O
may	O
result	O
in	O
an	O
error	O
rate	O
of	O
5%	O
(	O
95%	O
accuracy	O
)	O
or	O
worse	O
if	O
the	O
measurement	O
is	O
based	O
on	O
whether	O
each	O
whole	O
word	O
was	O
recognized	O
with	O
no	O
incorrect	O
letters	O
.	O
</s>
<s>
Using	O
a	O
large	O
enough	O
dataset	O
is	O
so	O
important	O
in	O
a	O
neural	B-Architecture
network	I-Architecture
based	O
handwriting	B-Application
recognition	I-Application
solutions	O
.	O
</s>
<s>
Web-based	O
OCR	O
systems	O
for	O
recognizing	O
hand-printed	O
text	O
on	O
the	O
fly	O
have	O
become	O
well	O
known	O
as	O
commercial	O
products	O
in	O
recent	O
years	O
(	O
see	O
Tablet	B-Device
PC	I-Device
history	I-Device
)	O
.	O
</s>
