Face Verification Using Colour Kernels

In this paper, the problem of fusing colour information at the feature level using kernel based feature extraction techniques is considered within the framework of a face verification system. A few statistical feature extraction algorithms including kernel based methods are reviewed first. An automatic parameter selection process is then applied to optimise the adopted kernel methods. In an extensive experimentation on intensity images of the XM2VTS database, we show that the optimised nonlinear kernel methods in general and the GDA algorithm in particular outperform the basic linear approaches. The experimentation is repeated for colour images by concatenating the R,G,B vectors. We demonstrate that by combining the colour information using the proposed method, the performance of the resulting decision making scheme considerably improves.