Age and Gender Estimation Using Shifting and Re-scaling of Local Regions - Technologies and Applications of Artificial Intelligence - page 4

Fig. . 1. The flow of our proposed algorithm

For added robustness, w
we slide each window by a
from each region of interes
ture describing the region o
To capture the largest am
multiple times and apply th
cale each window to 0.8 of
window to 1.2 of its origin
levels of details from the im
Once we have features r
ther gender or age), we use

we add a sliding window for each region of interest wh
a certain amount. This way we get 9 windows or patc
st, and we take the average of their histograms as the f
of interest.
mount of details possible, we rescale each region of inte
he LBP operator once on each scale. For example, we r
f its original size and apply LBP on it, then we rescale
nal size and apply LBP. This allows us to capture differ
mage.
ready and paired up with the proper label (representing
SVM for training and testing.

here
ches
fea-

erest
res-
the
rent

g ei-

4.2 Regions of Interest( (ROI) Selection

There is a huge variation b
such as the alignment of fa
main facial features. To se
acquired by the supervised

etween faces. To solve the challenges that come with th
acial features, we choose to focus on local regions arou
elect these regions accurately we use the landmark po
descent method.

that,
und
oints

pointed on a face [18]. (b) The same facial landmarks numbere Fig. 2. (a) Facial landmarks p ed.