Information Technology Reference
In-Depth Information
machine learning based approach where a cascade function is trained with negative
and positive samples and then could be used to detect objects in other images.
3.2
Landmark Extraction Using Supervised Descent Method
We will describe our method to detect the landmark points on the face, which is Su-
pervised Descent Method (SDM) and its Applications to Face Alignment [23]. The
main difference between SDM and other gradient descent methods is that during
training, the SDM learns a sequence of descent directions that minimize the mean of
Non-linear Least Squares (NLS) functions sampled at different points. Then in test-
ing, the SDM minimizes the NLS objective using the learned descent directions.
The training begins with finding the face in the training images, and then the initial
shape estimate is given by centering the mean face at the square. Then the translation-
al and scaling differences between the initial and manually labeled landmarks are
computed. Then Scale Invariant Feature Transform (SIFT) [21] descriptors are com-
puted on 32 x 32 pixel local patches. The SIFT features offer robust representation
against illumination changes. The goal is to learn a sequence of descent directions and
re-scaling factors such that it produces a sequence of updates starting from the initial
estimate that converges to the ground truth location.
3.3
Rotation Invariant Local Binary Pattern
Local Binary Pattern (LBP) is a texture operator that was first introduced by Ojala et
al. [19] based on the assumption that texture has locally two complementary aspects:
a pattern and strength. One of the most attractive features of LBP is its robustness to
gray-scale changes caused by differences in illumination.
3.4
Support Vector Machines
Support Vector Machines (SVMs)[1][4] are a useful technique for data classification
and regression. Given a set of training samples each labeled with a class label and a
set of features. The goal is to find a model to predict the label for the test data given
only the features for the test data.
4
Methodology
4.1
Overview
Our age and gender estimation system uses frontal face images for training and test-
ing. The first step is to detect whether or not a face exists and find its location using
the Haar cascades. After that, we find the cranio-facial landmarks on the face using
the supervised gradient descent method, so we can separate the face into different
regions. After separating the face into different regions of interest (left eyebrow, right
eyebrow, left eye, right eye, nose, and mouth), we begin to extract features, the fea-
tures we choose are the histogram of the rotation-invariant local binary pattern.
Search WWH ::
Custom Search