A novel approach based on variance for local feature analysis of facial images

A low dimensional representation of sensory signals is a key for solving many of the computational problems encountered in high level vision. In this paper, a comparison of face recognition techniques using principal component analysis (PCA) is done with local feature analysis (LFA) and an alternate method based on variance for quickly finding the local feature points on face images is also proposed. The LFA method is an extension of the eigenfaces method and gives a low-dimensional output for face representation. Principal component analysis (PCA) that is used for dimensionality reduction in the eigenfaces technique leads to global outputs, which are non-topographic and are not biologically plausible. On the other hand, the local feature analysis (LFA) technique yields local, topographic outputs which are sparsely distributed. They are effectively low dimensional but retain all the characteristics of the global modes. Local representations are desirable since they offer robustness against variability due to changes in the localised regions of the objects. A strategy for recognising faces using LFA is also proposed and several results on reconstruction and recognition are given to compare the performance of the variance method with that of LFA and PCA.