Nonlinear features for improved pattern recognition

Nonlinear features that represent higher-order correlations in input data are considered for improved recognition. They optimize new performance measures that do not make Gaussian etc. data distribution assumptions and that are intended for improved discrimination. The new features are produced in closed-form and are thus preferable to iterative solutions. An efficient two-step feature extraction algorithm is presented for the high-dimensional (iconic) input data case of most interest. The feature generation can be realized as a new neural network with adaptive activation functions. Test results on pose-invariant face recognition are emphasized; results on standard feature inputs for a product inspection application are briefly noted as a low-dimensional input data case.