A local face statistics recognition methodology beyond ICA and/or PCA

We have reviewed the independent component analysis (ICA), as an unsupervised ANN learning algorithm for redundancy reduction and feature extraction, and compared its performance with the classical principal component analysis (PCA) of face images, known as "eigenfaces". Based on our experiments, we believe that with PCA and ICA representations, a promising 85% to 95% PD with approximately 5% to 10% FAR in the ROC experiments might be achieved for a closed library set of persons, each of which has different profiles and lightening expressions. ICA encodes face images with statistically independent variables, which are not necessarily associated with the orthogonal axes, while PCA is always associated with orthogonal eigenvectors. Sometimes, the projections onto the ICA non-orthogonal axes are above the recognition threshold while the projections upon the orthogonal PCA axes are under the threshold However, both these pixel-based statistical processing algorithms have their drawbacks. The major one is that they weight the whole face equally and therefore lack the local geometry information. We argue that a fully robust face recognition or pattern recognition system should take both the gestalt geometry principle and the individual statistical features into account, i.e. it should approach from both statistical and geometry perspectives. An efficient way to implement both is the local or regional statistics, which may be called the local ICA or local PCA.