<s>
VisualRank	B-Algorithm
is	O
a	O
system	O
for	O
finding	B-General_Concept
and	O
ranking	O
images	O
by	O
analysing	O
and	O
comparing	O
their	O
content	O
,	O
rather	O
than	O
searching	O
image	O
names	O
,	O
Web	O
links	O
or	O
other	O
text	O
.	O
</s>
<s>
Google	B-Application
scientists	O
made	O
their	O
VisualRank	B-Algorithm
work	O
public	O
in	O
a	O
paper	O
describing	O
applying	O
PageRank	B-Algorithm
to	O
Google	B-Application
image	B-General_Concept
search	I-General_Concept
at	O
the	O
International	O
World	O
Wide	O
Web	O
Conference	O
in	O
Beijing	O
in	O
2008	O
.	O
</s>
<s>
Both	O
computer	B-Application
vision	I-Application
techniques	O
and	O
locality-sensitive	B-Algorithm
hashing	I-Algorithm
(	O
LSH	O
)	O
are	O
used	O
in	O
the	O
VisualRank	B-Algorithm
algorithm	O
.	O
</s>
<s>
Consider	O
an	O
image	B-General_Concept
search	I-General_Concept
initiated	O
by	O
a	O
text	O
query	O
.	O
</s>
<s>
An	O
existing	O
search	O
technique	O
based	O
on	O
image	O
metadata	O
and	O
surrounding	O
text	O
is	O
used	O
to	O
retrieve	O
the	O
initial	O
result	O
candidates	O
(	O
PageRank	B-Algorithm
)	O
,	O
which	O
along	O
with	O
other	O
images	O
in	O
the	O
index	O
are	O
clustered	O
in	O
a	O
graph	B-Application
according	O
to	O
their	O
similarity	O
(	O
which	O
is	O
precomputed	O
)	O
.	O
</s>
<s>
VisualRank	B-Algorithm
is	O
defined	O
iteratively	O
by	O
,	O
where	O
is	O
the	O
image	O
similarity	O
matrix	O
.	O
</s>
<s>
Clearly	O
,	O
the	O
image	O
similarity	O
measure	O
is	O
crucial	O
to	O
the	O
performance	O
of	O
VisualRank	B-Algorithm
since	O
it	O
determines	O
the	O
underlying	O
graph	B-Application
structure	O
.	O
</s>
<s>
The	O
main	O
VisualRank	B-Algorithm
system	O
begins	O
with	O
local	O
feature	O
vectors	O
being	O
extracted	O
from	O
images	O
using	O
scale-invariant	B-Algorithm
feature	I-Algorithm
transform	I-Algorithm
(	O
SIFT	O
)	O
.	O
</s>
<s>
Locality-sensitive	B-Algorithm
hashing	I-Algorithm
is	O
then	O
applied	O
to	O
these	O
feature	O
vectors	O
using	O
the	O
p-stable	O
distribution	O
scheme	O
.	O
</s>
