Information Technology Reference
In-Depth Information
Fig. 2.
Example images that are not considered
Fig. 3.
The pre-processing process on face images. (
left
) original image; (
middle
)
aligned image; (
right
)croppedface.
2.2 Our Approach
A gender recognition system consists of two key components: facial feature ex-
traction and classifier design. As reviewed in Section 1 and Table 1, raw image
pixels and Haar-like features are two often-used representations. In this work, we
employ LBP features, which have been widely exploited for facial representation
in recent years [13]. The most important properties of LBP features are their tol-
erance against monotonic illumination changes and their computational simplic-
ity. The LBP operator labels the image pixels by thresholding a neighborhood of
each pixel with the center value and considering the results as a binary number.
As shown in Fig. 4, face images are divided into non-overlapping sub-regions,
and the LBP histograms extracted from each sub-region are concatenated into
a feature histogram. Following the parameter settings suggested in [13], in our
experiments, face images of 127
91 pixels were divided into 42 sub-regions of
×
18
15 pixels, and the 59-label
LBP
(8
,
2
,u
2) operator [13] was adopted to ex-
tract LBP features. Thus each face image was described by a LBP histogram
of 2,478 (42
×
59) bins. With the LBP-based representation, SVM can be used
for gender classification, which has been an effective classification method in
existing studies [5,9].
Beyond the above standard representation, we further adopt Adaboost to
learn the discriminative LBP-Histogram (LBPH) bins for gender classification.
Adaboost has proved effective in both accuracy and speed for gender classifica-
tion [10,8]. Here we aim to select the LBPH bins which best separate the female
and male samples. The weak classifier
h
j
(
x
) consists of a feature
f
j
which corre-
sponds to the LBPH bin, a threshold
θ
j
and a parity
p
j
indicating the direction
of the inequality sign:
×
h
j
(
x
)=
1if
p
j
f
j
(
x
)
p
j
θ
j
≤
(1)
0otherwise
Search WWH ::
Custom Search