Information Technology Reference
In-Depth Information
desired. Ideally we would have the face of each person throughout his whole life; at
least one picture per year, but most databases have pictures of people at a certain age
only. Moreover, the images at higher ages are especially rare.
The goal of this paper is to implement an age and gender estimation algorithm us-
ing frontal facing facial images where the input would be an image and the output
would be estimation for the age and the gender, respectively.
2
Related Works
There are many aging face models. The models recognized in [6] are:
•
Anthropometric Model.
•
Active Appearance Model.
•
Aging Pattern Subspace.
•
Age Manifold.
•
Appearance Model.
Once we get our aging feature representation, the next step would be to estimate
the age. The choice of your aging feature representation should affect our choice for
age estimation algorithm, for example if we had strong facial features our age estima-
tion algorithm should be simple and if we choose simple facial features our age esti-
mation algorithm should be complex.
While gender classification is clearly a classification problem (Only two classes to
choose from; male and female). The question that begs itself is whether age estima-
tion is a classification problem or a regression problem. Age estimation can be looked
at as a pattern recognition task where each age label can be viewed as a class; there-
for, age estimation can be looked at as a classification problem. Another way to look
at the problem is that age numbers are a set of sequential numbers; therefore, age
estimation can be considered as a regression problem.
Thus age estimation can be viewed as a classification or a regression problem. To
find out which is more suitable, we need to try both of them on the same databases to
see the performance difference. Guo et al. [10, 11] did that and they chose Support
Vector Machine (SVM) as the classifier and Support Vector Regression (SVR) as the
regressor. They used the same image database for both experiments to evaluate their
performance. The result depended on which database they used, on the YGA database
the SVM outperformed the SVR. While with the FG-NET [5] database, it was the
other way around. This leads to suggest that it might be wise to combine classification
and regression for higher robustness and to get benefit from both methods.
3
Background
3.1
Face Detection Using Haar Cascades
To estimate the age and gender, we need to find the location of the face first. One of
the methods to detect faces is using Haar Feature-based cascade classifiers [24]. It is a
Search WWH ::
Custom Search