Information Technology Reference
In-Depth Information
Fig. 3 Front End: Face Detection and Feature Extraction
made. If this process is applied iteratively, the training error of the strong classifier
exponentially approaches zero. This is repeated until the error rate drops below a set
threshold.
Viola-Jones uses Adaboost both for feature selection and to train the face detec-
tor. During feature selection, Adaboost can select a few important features from the
set of all possible features. The features are Haar-like wavelets. The Haar-like fea-
tures are extracted by applying a threshold to the sums and differences of rectangular
image regions. O PEN CV also includes diagonal wavelets 4 to detect rectangular fea-
tures rotated 45 . Each weak classifier in O PEN CV typically uses one feature (at
most three).
A weak classifier h
(
x
,
f
,
p
, θ )
is used, where x is a 24
×
24 pixel sub-window
of an image, f is a feature,
θ
is a threshold and p is the polarity (direction) of the
inequality:
1if pf
(
x
) <
p
θ
h
(
x
,
f
,
p
, θ )=
0otherwise
Each stage of the boosting process attempts to find the Haar-like feature which
best separates faces from non-faces.
One of Viola-Jones' innovations was the use of integral images to rapidly com-
pute the Haar-like features at any scale or location in the image. The integral image
at
(
,
)
(
,
)
.Thesum
of all pixels within any given rectangle can be rapidly calculated from the sums and
differences of the four corner pixels of the integral image. By thresholding these
rectangles, the Haar-like features can be calculated at any location or scale without
needing to build a multi-scale pyramid.
x
y
is defined as the sum of all pixels above and to the left of
x
y
4
This extension to the Viola-Jones technique is described in [16].
 
Search WWH ::




Custom Search