Fast and efficient feature extraction based on Bayesian decision boundaries

The implementation of a pattern recognition system requires solutions to some basic problems: data acquisition, feature extraction and pattern classification. In this paper a novel and efficient approaches for feature extraction for pattern classification using neural networks is proposed. The method searches for the minimum amount of features necessary for solving a given pattern classification problem based on the structure of an adequately trained MLP network. Experimentally we show that all informative discriminating features can be obtained from decision boundaries specified by the MLP network