Feature extraction of facial action units combining kernel methods and independent component analysis

Abstract

The work described in this paper presents a methodology for characterizing facial action units (AUs), which represents the subtle change of facial expressions, based on Kernels Methods perform a nonlinear maping of data and looking for directions to the projections of the data in feature space through independent component analysis (ICA). The methodology validation was done on Cohn-Kande database. Image preprocessing was done through histogram equalization, a whitening on data with Kernel Principal Component Analysis (KPCA), for that the mapped in feature space search a lineal structure of the input data, finally we applied ICA for make the projected distribution of data is at least possible Gaussian. The results were 96.64% accurate for average recognition of three combinations of facial AUs of the whole face more neutral faces. Mainly changes that occur between rapid transitions of AUs instantly shown were detected. Additionally the proposed methodology can reduce the size of the feature space because data only in terms of independent components, are presented so as to use only those variables that provided greater information, which reduces the complexity of the classifier.

Keywords

• facial expression recognition
• action unit
• independent component analysis
• kernel methods