PDF version of this entire document

A PCA Approach

The idea adopted here is one which involves learning variation from a large dataset. Principal component analysis (PCA) gets used for this and along the lines of Cootes et al. [,,], biologically-meaningful modes of variation are found which encompass a set of faces (albeit in 3-D rather than 2-D). Once the average face is known - as extracted from the data along with the common modes of variation - it is then possible to apply transformations in reverse, reparameterising the model according to the problem domain. If all images can be brought into a common frame of reference, comparison is made trivial, using known similarity measures. It is important to only model facial expressions, however, excepting nose and forehead for instance. Viewing the outcome in terms of recognition rates (with http://en.wikipedia.org/wiki/Roc_curveROC curves for example) enables tweaking and fine-tuning the method.



Roy Schestowitz 2012-01-08