<s>
Haar-like	B-Algorithm
features	I-Algorithm
are	O
digital	B-Algorithm
image	I-Algorithm
features	O
used	O
in	O
object	O
recognition	O
.	O
</s>
<s>
Historically	O
,	O
working	O
with	O
only	O
image	O
intensities	O
(	O
i.e.	O
,	O
the	O
RGB	O
pixel	B-Algorithm
values	O
at	O
each	O
and	O
every	O
pixel	B-Algorithm
of	O
image	O
)	O
made	O
the	O
task	O
of	O
feature	B-Algorithm
calculation	O
computationally	B-General_Concept
expensive	I-General_Concept
.	O
</s>
<s>
discussed	O
working	O
with	O
an	O
alternate	O
feature	B-Algorithm
set	O
based	O
on	O
Haar	O
wavelets	O
instead	O
of	O
the	O
usual	O
image	O
intensities	O
.	O
</s>
<s>
Paul	O
Viola	O
and	O
Michael	O
Jones	O
adapted	O
the	O
idea	O
of	O
using	O
Haar	O
wavelets	O
and	O
developed	O
the	O
so-called	O
Haar-like	B-Algorithm
features	I-Algorithm
.	O
</s>
<s>
A	O
Haar-like	B-Algorithm
feature	I-Algorithm
considers	O
adjacent	O
rectangular	O
regions	O
at	O
a	O
specific	O
location	O
in	O
a	O
detection	O
window	O
,	O
sums	O
up	O
the	O
pixel	B-Algorithm
intensities	O
in	O
each	O
region	O
and	O
calculates	O
the	O
difference	O
between	O
these	O
sums	O
.	O
</s>
<s>
Therefore	O
,	O
a	O
common	O
Haar	O
feature	B-Algorithm
for	O
face	O
detection	O
is	O
a	O
set	O
of	O
two	O
adjacent	O
rectangles	O
that	O
lie	O
above	O
the	O
eye	O
and	O
the	O
cheek	O
region	O
.	O
</s>
<s>
In	O
the	O
detection	O
phase	O
of	O
the	O
Viola	B-General_Concept
–	I-General_Concept
Jones	I-General_Concept
object	I-General_Concept
detection	I-General_Concept
framework	I-General_Concept
,	O
a	O
window	O
of	O
the	O
target	O
size	O
is	O
moved	O
over	O
the	O
input	O
image	O
,	O
and	O
for	O
each	O
subsection	O
of	O
the	O
image	O
the	O
Haar-like	B-Algorithm
feature	I-Algorithm
is	O
calculated	O
.	O
</s>
<s>
Because	O
such	O
a	O
Haar-like	B-Algorithm
feature	I-Algorithm
is	O
only	O
a	O
weak	O
learner	O
or	O
classifier	O
(	O
its	O
detection	O
quality	O
is	O
slightly	O
better	O
than	O
random	O
guessing	O
)	O
a	O
large	O
number	O
of	O
Haar-like	B-Algorithm
features	I-Algorithm
are	O
necessary	O
to	O
describe	O
an	O
object	O
with	O
sufficient	O
accuracy	O
.	O
</s>
<s>
In	O
the	O
Viola	B-General_Concept
–	I-General_Concept
Jones	I-General_Concept
object	I-General_Concept
detection	I-General_Concept
framework	I-General_Concept
,	O
the	O
Haar-like	B-Algorithm
features	I-Algorithm
are	O
therefore	O
organized	O
in	O
something	O
called	O
a	O
classifier	O
cascade	O
to	O
form	O
a	O
strong	O
learner	O
or	O
classifier	O
.	O
</s>
<s>
The	O
key	O
advantage	O
of	O
a	O
Haar-like	B-Algorithm
feature	I-Algorithm
over	O
most	O
other	O
features	O
is	O
its	O
calculation	O
speed	O
.	O
</s>
<s>
Due	O
to	O
the	O
use	O
of	O
integral	B-Algorithm
images	I-Algorithm
,	O
a	O
Haar-like	B-Algorithm
feature	I-Algorithm
of	O
any	O
size	O
can	O
be	O
calculated	O
in	O
constant	O
time	O
(	O
approximately	O
60	O
microprocessor	O
instructions	O
for	O
a	O
2-rectangle	O
feature	B-Algorithm
)	O
.	O
</s>
<s>
A	O
simple	O
rectangular	O
Haar-like	B-Algorithm
feature	I-Algorithm
can	O
be	O
defined	O
as	O
the	O
difference	O
of	O
the	O
sum	O
of	O
pixels	B-Algorithm
of	O
areas	O
inside	O
the	O
rectangle	O
,	O
which	O
can	O
be	O
at	O
any	O
position	O
and	O
scale	O
within	O
the	O
original	O
image	O
.	O
</s>
<s>
This	O
modified	O
feature	B-Algorithm
set	O
is	O
called	O
2-rectangle	O
feature	B-Algorithm
.	O
</s>
<s>
Each	O
feature	B-Algorithm
type	O
can	O
indicate	O
the	O
existence	O
(	O
or	O
absence	O
)	O
of	O
certain	O
characteristics	O
in	O
the	O
image	O
,	O
such	O
as	O
edges	O
or	O
changes	O
in	O
texture	O
.	O
</s>
<s>
For	O
example	O
,	O
a	O
2-rectangle	O
feature	B-Algorithm
can	O
indicate	O
where	O
the	O
border	O
lies	O
between	O
a	O
dark	O
region	O
and	O
a	O
light	O
region	O
.	O
</s>
<s>
One	O
of	O
the	O
contributions	O
of	O
Viola	O
and	O
Jones	O
was	O
to	O
use	O
summed-area	B-Algorithm
tables	I-Algorithm
,	O
which	O
they	O
called	O
integral	B-Algorithm
images	I-Algorithm
.	O
</s>
<s>
Integral	B-Algorithm
images	I-Algorithm
can	O
be	O
defined	O
as	O
two-dimensional	O
lookup	B-Data_Structure
tables	I-Data_Structure
in	O
the	O
form	O
of	O
a	O
matrix	O
with	O
the	O
same	O
size	O
of	O
the	O
original	O
image	O
.	O
</s>
<s>
Each	O
element	O
of	O
the	O
integral	B-Algorithm
image	I-Algorithm
contains	O
the	O
sum	O
of	O
all	O
pixels	B-Algorithm
located	O
on	O
the	O
up-left	O
region	O
of	O
the	O
original	O
image	O
(	O
in	O
relation	O
to	O
the	O
element	O
's	O
position	O
)	O
.	O
</s>
<s>
This	O
allows	O
to	O
compute	O
sum	O
of	O
rectangular	O
areas	O
in	O
the	O
image	O
,	O
at	O
any	O
position	O
or	O
scale	O
,	O
using	O
only	O
four	O
lookups	B-Data_Structure
:	O
</s>
<s>
where	O
points	O
belong	O
to	O
the	O
integral	B-Algorithm
image	I-Algorithm
,	O
as	O
shown	O
in	O
the	O
figure	O
.	O
</s>
<s>
Each	O
Haar-like	B-Algorithm
feature	I-Algorithm
may	O
need	O
more	O
than	O
four	O
lookups	B-Data_Structure
,	O
depending	O
on	O
how	O
it	O
was	O
defined	O
.	O
</s>
<s>
Viola	O
and	O
Jones	O
's	O
2-rectangle	O
features	O
need	O
six	O
lookups	B-Data_Structure
,	O
3-rectangle	O
features	O
need	O
eight	O
lookups	B-Data_Structure
,	O
and	O
4-rectangle	O
features	O
need	O
nine	O
lookups	B-Data_Structure
.	O
</s>
<s>
Lienhart	O
and	O
Maydt	O
introduced	O
the	O
concept	O
of	O
a	O
tilted	O
(	O
45°	O
)	O
Haar-like	B-Algorithm
feature	I-Algorithm
.	O
</s>
<s>
For	O
example	O
,	O
a	O
2-rectangle	O
tilted	O
Haar-like	B-Algorithm
feature	I-Algorithm
can	O
indicate	O
the	O
existence	O
of	O
an	O
edge	O
at	O
45°	O
.	O
</s>
<s>
Messom	O
and	O
Barczak	O
extended	O
the	O
idea	O
to	O
a	O
generic	O
rotated	O
Haar-like	B-Algorithm
feature	I-Algorithm
.	O
</s>
<s>
Although	O
the	O
idea	O
is	O
sound	O
mathematically	O
,	O
practical	O
problems	O
prevent	O
the	O
use	O
of	O
Haar-like	B-Algorithm
features	I-Algorithm
at	O
any	O
angle	O
.	O
</s>
<s>
In	O
order	O
to	O
be	O
fast	O
,	O
detection	O
algorithms	O
use	O
low	O
resolution	O
images	O
introducing	O
rounding	B-Algorithm
errors	I-Algorithm
.	O
</s>
<s>
For	O
this	O
reason	O
rotated	O
Haar-like	B-Algorithm
features	I-Algorithm
are	O
not	O
commonly	O
used	O
.	O
</s>
