We wish to identify a fusion operator P_w : R~d -> R~k which reduces d-dimensional sensor imagery to the k dimensions that allow the best classification performance. We present a bilevel optimization procedure for maximizing, conditional on observed training data and subject to maximum likelihood constraints, a measure of the k-dimensional class-conditional probability density estimate separation for k < d. the separation optimization involves traversal of a manifold of maximum likelihood solutions parameterized by the projection vector w, the design variable. This optimization is performed in pursuit of the ultimate object6ive: minimizing (over w implied by W) the probability of misclassification when using a Bayes classifier based on k-dimensional maximum likelihood class-conditional mixture model density estimates. A penalty interior point approach to the required optimization is proposed which generates a solution that satisfies the fundamental first-order optimlity conditions. The performance of the proposed algorithm is illustrated through the application to sensor fusion in a multispectral minefield detection problem.
展开▼