<s>
Document	B-Algorithm
mosaicing	I-Algorithm
is	O
a	O
process	O
that	O
stitches	B-Algorithm
multiple	O
,	O
overlapping	O
snapshot	O
images	O
of	O
a	O
document	O
together	O
to	O
produce	O
one	O
large	O
,	O
high	O
resolution	O
composite	O
.	O
</s>
<s>
The	O
system	O
then	O
finds	O
the	O
overlapped	O
pairs	O
and	O
stitches	B-Algorithm
them	O
together	O
repeatedly	O
until	O
all	O
pairs	O
are	O
stitched	O
together	O
as	O
one	O
piece	O
of	O
document	O
.	O
</s>
<s>
The	O
document	B-Algorithm
mosaicing	I-Algorithm
can	O
be	O
divided	O
into	O
four	O
main	O
processes	O
.	O
</s>
<s>
Each	O
image	O
is	O
segmented	B-Algorithm
into	O
a	O
hierarchy	O
of	O
columns	O
,	O
lines	O
,	O
and	O
words	O
to	O
match	O
the	O
organised	O
sets	O
of	O
features	O
across	O
images	O
.	O
</s>
<s>
Firstly	O
,	O
the	O
angle	O
that	O
the	O
rows	O
of	O
text	O
make	O
with	O
the	O
image	O
raster	B-Algorithm
lines	O
(	O
skew	O
angle	O
)	O
is	O
estimated	O
.	O
</s>
<s>
A	O
small	O
patch	O
of	O
text	O
in	O
the	O
image	O
is	O
selected	O
randomly	O
and	O
then	O
rotated	O
in	O
the	O
range	O
of	O
±20°	O
until	O
the	O
variance	O
of	O
the	O
pixel	O
intensities	O
of	O
the	O
patch	O
summed	O
along	O
the	O
raster	B-Algorithm
lines	O
is	O
maximised	O
.	O
</s>
<s>
In	O
this	O
operation	O
,	O
the	O
de-skewed	O
document	O
is	O
intuitively	O
segmented	B-Algorithm
into	O
a	O
hierarchy	O
of	O
columns	O
,	O
lines	O
and	O
words	O
.	O
</s>
<s>
The	O
sensitivity	O
to	O
illumination	O
and	O
page	O
coloration	O
of	O
the	O
de-skewed	O
document	O
can	O
be	O
removed	O
by	O
applying	O
a	O
Sobel	B-Algorithm
operator	I-Algorithm
to	O
the	O
de-skewed	O
image	O
and	O
thresholding	O
the	O
output	O
to	O
obtain	O
the	O
binary	O
gradient	O
,	O
de-skewed	O
image	O
.	O
</s>
<s>
The	O
operation	O
can	O
be	O
roughly	O
separated	O
into	O
3	O
steps	O
:	O
column	O
segmentation	B-Algorithm
,	O
line	O
segmentation	B-Algorithm
and	O
word	O
segmentation	B-Algorithm
.	O
</s>
<s>
Columns	O
are	O
easily	O
segmented	B-Algorithm
from	O
the	O
binary	O
gradient	O
,	O
de-skewed	O
images	O
by	O
summing	O
pixels	O
vertically	O
as	O
shown	O
in	O
Figure	O
4	O
.	O
</s>
<s>
Baselines	O
of	O
each	O
row	O
are	O
segmented	B-Algorithm
in	O
the	O
same	O
way	O
as	O
the	O
column	O
segmentation	B-Algorithm
process	O
but	O
horizontally	O
.	O
</s>
<s>
Finally	O
,	O
individual	O
words	O
are	O
segmented	B-Algorithm
by	O
applying	O
the	O
vertical	O
process	O
at	O
each	O
segmented	B-Algorithm
row	O
.	O
</s>
<s>
These	O
segmentations	B-Algorithm
are	O
important	O
because	O
the	O
document	O
mosaic	O
is	O
created	O
by	O
matching	O
the	O
lower	O
right	O
corners	O
of	O
words	O
in	O
overlapping	O
images	O
pair	O
.	O
</s>
<s>
Moreover	O
,	O
the	O
segmentation	B-Algorithm
operation	O
can	O
organize	O
the	O
list	O
of	O
images	O
in	O
the	O
context	O
of	O
a	O
hierarchy	O
of	O
rows	O
and	O
column	O
reliably	O
.	O
</s>
<s>
Assuming	O
a	O
pinhole	B-Algorithm
camera	I-Algorithm
model	O
,	O
the	O
transformation	O
between	O
pixels	O
(	O
u	O
,	O
v	O
)	O
of	O
image	O
1	O
and	O
pixels	O
(	O
u0	O
,	O
v0	O
)	O
of	O
image	O
2	O
is	O
demonstrated	O
by	O
a	O
plane-to-plane	O
projectivity	O
.	O
</s>
<s>
There	O
are	O
various	O
areas	O
that	O
the	O
technique	O
of	O
document	B-Algorithm
mosaicing	I-Algorithm
can	O
be	O
applied	O
to	O
such	O
as	O
:	O
</s>
