Feature Extraction by Foley-Sammon Transform with Kernels

A method KFST (Foley-Sammon transform with kernels)is proposed which is based on FST (Foley-Sammon transform) and kernel tricks. The projectors onto the directions derived by KFST can be used for class-specific feature extraction. The algorithm is carried out in a feature space associated with kernel functions, hence it can be used to construct a large class of nonlinear feature extractors. Linear feature extraction in feature space corresponds to nonlinear feature extraction in input space. KFST is proven to correspond to a generalized eigenvalue problem. Lastly, our method is applied to digits and images recognition problems, and the experimental results show that present method is superior to the existing methods in term of space distribution and correct classification rate.