Evolution of optimal projection axes (OPA) for face recognition

The paper describes a novel approach called Optimal Projection Axes (OPA) for face recognition. OPA works by searching through all the rotations defined over whitened principal component analysis (PCA) subspaces. Whitening, which does not preserve norms, plays a dual role: (i) counteracts the fact that the mean square error (MSE) principle underlying PCA preferentially weights low frequencies; and (ii) increases the reachable space of solutions to include non orthogonal bases. Better performance from non orthogonal bases over orthogonal ones is expected as they lead to an overcomplete and robust representational space. As the search space is too large for any systematic search, stochastic and directed ("greedy") search is undertaken using evolution in the form of genetic algorithms (GAs). Evolution is driven by a fitness function defined in terms of performance accuracy and class separation (scatter index). Accuracy indicates the extent to which learning has been successful so far while the scatter index gives an indication of the expected fitness on future trials. Experiments carried out using 1107 facial images corresponding to 369 subjects (with 169 subjects having duplicated images) from the FERET database show that OPA yields improved performance over the eigenface and MDF (Most Discriminant Features) methods.