Kernel-based Global Folley-Sammon Discriminate Analysis

Kernel trick is a useful method dealing with the nonlinearity of patterns \´ distribution, and widely used for non-linear pattern classification problem. Recently, many kernel-based Fisher discriminate methods were proposed, the features extracted by those methods are statistically uncorrelated, usually ineffectively when used in so called "small sample problem "(SSS), which exists in most FR tasks. To overcome this shortcoming, in this paper we present Kernel-based Global Folley-Sammon Discriminate Analysis (KGFS), it can be realized in three steps: firstly by combing kernel function, map data onto an implicit high-dimensional feature space, secondly construct orthogonal feature space there, finally calculate iterative global orthogonal discriminate vectors. Experiments have been done on ORL dataset. In terms of classification error rate performance, proposed kernel based method KGFS has a better performance than other three commonly used kernel-based methods, such as GDA, MGDAandKDDA.