Expression Classification Using Wavelet Asymmetry Faces
Fall 2005, Carnegie Mellon University

I investigated wavelet packet method on asymmetry faces for expression classification. Two types of quantified asymmetry measures have already been defined, but they are computed only in the image intensity domain. I am proposing a novel approach to extend those asymmetry faces by using wavelet transforms to achieve better and more robust results. One of the most interesting characteristics of wavelet transforms is their ability to represent a signal into partitions of time-frequency plane. Using a random subset of 3 expressions with 55 subjects each from a normalized version of the Cohn-Kanade AU-Coded Facial Expression Database, error reduction rates of 25-95% have been achieved. My findings show that asymmetry faces have some features that remain more consistent and discriminative in the wavelet subspaces than in the image intensity domain. I also show that certain subspaces will contribute more than others in classification problems.

Proceed to details