Age and Gender Estimation Using Shifting and Re-scaling of Local Regions - Technologies and Applications of Artificial Intelligence - page 7

Fig. 5. (a) Face image of a man with the nose region highlighted [18]. (b) The 9 cropped areas
taken by sliding the original ROI by s pixels (9 pixels in this example).

We start with the original position of the patch and move it to the left by s pixels
and upwards by s pixels. Then we increment the x and y position by s pixels indepen-
dently three times. This way, we have 9 patches around the original one. This should
solve any alignment issues in the local scale.
Before we apply the LBP operator, we rescale each patch multiple times (in our
experiments, we take 3 sizes). We use bilinear interpolation to resize each patch.
Once we have the resized patches ready we can apply the invariant LBP operator.

4.4 SVM Training and Region Voting

Once we have each shifted and scaled patch ready, we apply the rotation-invariant LBP.
We take the histogram of the rotation-invariant LBP and average it out across each re-
gion and normalize it to its size. This will be our feature vector describing each region.
The next step is to train using the SVM for gender and SVR for age. We use
LIBSVM [2] for the implementation of the SVM and SVR. Each region would have
its own classifier and its own decision. The final decision (estimation) would be de-
cided by a majority vote where each region would cast its vote (male or female in
case of gender classification). Each region has an equally weighted vote, and since we
are using 7 regions, there would be no ties.