Research
on computer vision has been directed towards facial analysis in recent decades
and has actively gathered scientists from a variety of research perspectives. The
report presents our current efforts in facial features classification problems.
The pattern recognition problems that we have dealt with relate to gender, age,
expression, ethnicity and other facial properties. The report provides a
summary on the face dataset provided. The noisy real-world dataset was
characterized by uneven distributions of classes. However, the report
summarizes various feature selection techniques that seem to successfully
distinguish between classes. Localization of features shows its importance in
selecting image regions that seem to characterize the data in an unsupervised
manner that turns out to be more effective and intuitive than some of the
dimensionality reduction algorithms. The rest of the report describes some of
the classification algorithms used and their evaluation on the test dataset.
Feature
selection seems to be a non-trivial problem and is one of the many directions
to go. The quality of a feature set determines the success of a classification
algorithm. Correct choice of classifier
parameters also poses research challenges. We have not dealt with classifier
combinations in our current research. We will be studying the effect of
classifier combinations or some sort of mixture of experts in near future. We
also wish to use classification techniques like Regularized Least Squares
Classification, Bayes Point Machines, and so on. Various forms of graphical
models besides the ones used may turn out useful for the above task.
We
acknowledge Thorsten Joachims for SVMLight, Ryan Rifkin for multi class SVMFu,
Cognitive Machines for one-dimensional CDHMM implementations and D. Pantazis
for Intel OpenCV based face recognition
code using embedded hmms