<s>
Content-based	B-Application
image	I-Application
retrieval	I-Application
,	O
also	O
known	O
as	O
query	B-Application
by	I-Application
image	I-Application
content	I-Application
(	O
QBIC	B-Application
)	O
and	O
content-based	B-Application
visual	I-Application
information	I-Application
retrieval	I-Application
(	O
CBVIR	B-Application
)	O
,	O
is	O
the	O
application	O
of	O
computer	B-Application
vision	I-Application
techniques	O
to	O
the	O
image	B-General_Concept
retrieval	I-General_Concept
problem	O
,	O
that	O
is	O
,	O
the	O
problem	O
of	O
searching	O
for	O
digital	B-Algorithm
images	I-Algorithm
in	O
large	O
databases	O
(	O
see	O
this	O
survey	O
for	O
a	O
scientific	O
overview	O
of	O
the	O
CBIR	B-Application
field	O
)	O
.	O
</s>
<s>
Content-based	B-Application
image	I-Application
retrieval	I-Application
is	O
opposed	O
to	O
traditional	O
concept-based	O
approaches	O
(	O
see	O
Concept-based	O
image	O
indexing	O
)	O
.	O
</s>
<s>
CBIR	B-Application
is	O
desirable	O
because	O
searches	O
that	O
rely	O
purely	O
on	O
metadata	O
are	O
dependent	O
on	O
annotation	B-Application
quality	O
and	O
completeness	O
.	O
</s>
<s>
An	O
image	B-Application
meta	I-Application
search	I-Application
requires	O
humans	O
to	O
have	O
manually	O
annotated	O
images	O
by	O
entering	O
keywords	O
or	O
metadata	O
in	O
a	O
large	O
database	O
,	O
which	O
can	O
be	O
time-consuming	O
and	O
may	O
not	O
capture	O
the	O
keywords	O
desired	O
to	O
describe	O
the	O
image	O
.	O
</s>
<s>
The	O
evaluation	O
of	O
the	O
effectiveness	O
of	O
keyword	O
image	B-General_Concept
search	I-General_Concept
is	O
subjective	O
and	O
has	O
not	O
been	O
well-defined	O
.	O
</s>
<s>
In	O
the	O
same	O
regard	O
,	O
CBIR	B-Application
systems	O
have	O
similar	O
challenges	O
in	O
defining	O
success	O
.	O
</s>
<s>
The	O
term	O
"	O
content-based	B-Application
image	I-Application
retrieval	I-Application
"	O
seems	O
to	O
have	O
originated	O
in	O
1992	O
when	O
it	O
was	O
used	O
by	O
Japanese	O
Electrotechnical	O
Laboratory	O
engineer	O
Toshikazu	O
Kato	O
to	O
describe	O
experiments	O
into	O
automatic	O
retrieval	O
of	O
images	O
from	O
a	O
database	O
,	O
based	O
on	O
the	O
colors	O
and	O
shapes	O
present	O
.	O
</s>
<s>
The	O
techniques	O
,	O
tools	O
,	O
and	O
algorithms	O
that	O
are	O
used	O
originate	O
from	O
fields	O
such	O
as	O
statistics	O
,	O
pattern	O
recognition	O
,	O
signal	O
processing	O
,	O
and	O
computer	B-Application
vision	I-Application
.	O
</s>
<s>
The	O
earliest	O
commercial	O
CBIR	B-Application
system	O
was	O
developed	O
by	O
IBM	O
and	O
was	O
called	O
QBIC	B-Application
(	O
Query	B-Application
By	I-Application
Image	I-Application
Content	I-Application
)	O
.	O
</s>
<s>
While	O
the	O
storing	O
of	O
multiple	O
images	O
as	O
part	O
of	O
a	O
single	O
entity	O
preceded	O
the	O
term	O
BLOB	B-General_Concept
(	O
Binary	B-General_Concept
Large	I-General_Concept
OBject	I-General_Concept
)	O
,	O
the	O
ability	O
to	O
fully	O
search	O
by	O
content	O
,	O
rather	O
than	O
by	O
description	O
,	O
had	O
to	O
await	O
IBM	O
's	O
QBIC	B-Application
.	O
</s>
<s>
The	O
interest	O
in	O
CBIR	B-Application
has	O
grown	O
because	O
of	O
the	O
limitations	O
inherent	O
in	O
metadata-based	O
systems	O
,	O
as	O
well	O
as	O
the	O
large	O
range	O
of	O
possible	O
uses	O
for	O
efficient	O
image	B-General_Concept
retrieval	I-General_Concept
.	O
</s>
<s>
Initial	O
CBIR	B-Application
systems	O
were	O
developed	O
to	O
search	O
databases	O
based	O
on	O
image	O
color	O
,	O
texture	O
,	O
and	O
shape	O
properties	O
.	O
</s>
<s>
Therefore	O
,	O
efforts	O
in	O
the	O
CBIR	B-Application
field	O
started	O
to	O
include	O
human-centered	O
design	O
that	O
tried	O
to	O
meet	O
the	O
needs	O
of	O
the	O
user	O
performing	O
the	O
search	O
.	O
</s>
<s>
Many	O
CBIR	B-Application
systems	O
have	O
been	O
developed	O
,	O
but	O
,	O
the	O
problem	O
of	O
retrieving	O
images	O
on	O
the	O
basis	O
of	O
their	O
pixel	O
content	O
remains	O
largely	O
unsolved	O
.	O
</s>
<s>
Different	O
query	O
techniques	O
and	O
implementations	O
of	O
CBIR	B-Application
make	O
use	O
of	O
different	O
types	O
of	O
user	O
queries	O
.	O
</s>
<s>
QBE	O
(	O
Query	B-Application
By	I-Application
Example	I-Application
)	O
is	O
a	O
query	O
technique	O
that	O
involves	O
providing	O
the	O
CBIR	B-Application
system	O
with	O
an	O
example	O
image	O
that	O
it	O
will	O
then	O
base	O
its	O
search	O
upon	O
.	O
</s>
<s>
The	O
user	O
draws	O
a	O
rough	O
approximation	O
of	O
the	O
image	O
they	O
are	O
looking	O
for	O
,	O
for	O
example	O
with	O
blobs	B-General_Concept
of	O
color	O
or	O
general	O
shapes	O
.	O
</s>
<s>
Many	O
CBIR	B-Application
systems	O
therefore	O
generally	O
make	O
use	O
of	O
lower-level	O
features	O
like	O
texture	O
,	O
color	O
,	O
and	O
shape	O
.	O
</s>
<s>
However	O
,	O
in	O
general	O
,	O
image	B-General_Concept
retrieval	I-General_Concept
requires	O
human	O
feedback	O
in	O
order	O
to	O
identify	O
higher-level	O
concepts	O
.	O
</s>
<s>
Combining	O
CBIR	B-Application
search	O
techniques	O
available	O
with	O
the	O
wide	O
range	O
of	O
potential	O
users	O
and	O
their	O
intent	O
can	O
be	O
a	O
difficult	O
task	O
.	O
</s>
<s>
An	O
aspect	O
of	O
making	O
CBIR	B-Application
successful	O
relies	O
entirely	O
on	O
the	O
ability	O
to	O
understand	O
the	O
user	O
intent	O
.	O
</s>
<s>
CBIR	B-Application
systems	O
can	O
make	O
use	O
of	O
relevance	B-Application
feedback	I-Application
,	O
where	O
the	O
user	O
progressively	O
refines	O
the	O
search	O
results	O
by	O
marking	O
images	O
in	O
the	O
results	O
as	O
"	O
relevant	O
"	O
,	O
"	O
not	O
relevant	O
"	O
,	O
or	O
"	O
neutral	O
"	O
to	O
the	O
search	O
query	O
,	O
then	O
repeating	O
the	O
search	O
with	O
the	O
new	O
information	O
.	O
</s>
<s>
Machine	O
learning	O
and	O
application	O
of	O
iterative	O
techniques	O
are	O
becoming	O
more	O
common	O
in	O
CBIR	B-Application
.	O
</s>
<s>
Other	O
query	O
methods	O
include	O
browsing	O
for	O
example	O
images	O
,	O
navigating	O
customized/hierarchical	O
categories	O
,	O
querying	O
by	O
image	B-Algorithm
region	I-Algorithm
(	O
rather	O
than	O
the	O
entire	O
image	O
)	O
,	O
querying	O
by	O
multiple	O
example	O
images	O
,	O
querying	O
by	O
visual	O
sketch	O
,	O
querying	O
by	O
direct	O
specification	O
of	O
image	O
features	O
,	O
and	O
multimodal	B-Application
queries	O
(	O
e.g.	O
</s>
<s>
The	O
most	O
common	O
method	O
for	O
comparing	O
two	O
images	O
in	O
content-based	B-Application
image	I-Application
retrieval	I-Application
(	O
typically	O
an	O
example	O
image	O
and	O
an	O
image	O
from	O
the	O
database	O
)	O
is	O
using	O
an	O
image	O
distance	O
measure	O
.	O
</s>
<s>
Computing	O
distance	O
measures	O
based	O
on	O
color	O
similarity	O
is	O
achieved	O
by	O
computing	O
a	O
color	B-General_Concept
histogram	I-General_Concept
for	O
each	O
image	O
that	O
identifies	O
the	O
proportion	O
of	O
pixels	O
within	O
an	O
image	O
holding	O
specific	O
values	O
.	O
</s>
<s>
Shapes	O
will	O
often	O
be	O
determined	O
first	O
applying	O
segmentation	B-Algorithm
or	O
edge	B-Algorithm
detection	I-Algorithm
to	O
an	O
image	O
.	O
</s>
<s>
Like	O
other	O
tasks	O
in	O
computer	B-Application
vision	I-Application
such	O
as	O
recognition	O
and	O
detection	O
,	O
recent	O
neural	O
network	O
based	O
retrieval	O
algorithms	O
are	O
susceptible	O
to	O
adversarial	B-Algorithm
attacks	I-Algorithm
,	O
both	O
as	O
candidate	O
and	O
the	O
query	O
attacks	O
.	O
</s>
<s>
In	O
addition	O
,	O
model-agnostic	O
transferable	O
adversarial	O
examples	O
are	O
also	O
possible	O
,	O
which	O
enables	O
black-box	O
adversarial	B-Algorithm
attacks	I-Algorithm
on	O
deep	O
ranking	O
systems	O
without	O
requiring	O
access	O
to	O
their	O
underlying	O
implementations	O
.	O
</s>
<s>
Measures	O
of	O
image	B-General_Concept
retrieval	I-General_Concept
can	O
be	O
defined	O
in	O
terms	O
of	O
precision	O
and	O
recall	O
.	O
</s>
<s>
An	O
image	O
is	O
retrieved	O
in	O
CBIR	B-Application
system	O
by	O
adopting	O
several	O
techniques	O
simultaneously	O
such	O
as	O
Integrating	O
Pixel	O
Cluster	O
Indexing	O
,	O
histogram	O
intersection	O
and	O
discrete	O
wavelet	B-Algorithm
transform	I-Algorithm
methods	O
.	O
</s>
<s>
Potential	O
uses	O
for	O
CBIR	B-Application
include	O
:	O
</s>
