Characterization of Multilayer Perceptrons and Their Application to Multisensor Automatic Target Detection.

摘要

The multilayer perceptron was extensively analyzed. A technique for analyzing the multilayer perceptron, the saliency measure, was developed which provides a measure of the importance of inputs. The method was compared to the conventional statistical technique of best features and shown to provide similar rankings of the input. Using the saliency measure, it is shown that the multilayer perceptron effectively ignores useless inputs and that whether it is trained using backpropagation or extended Kalman filtering, the weighting of the inputs is the same. The backpropagation training algorithm is shown to be a degenerate version of the extended Kalman filter. The extended Kalman algorithm is shown to outperform the backpropagation method in terms of classification accuracy versus training presentations; however, in terms of computational complexity, the backpropagation algorithm is shown is shown to be highly efficient. The multilayer perceptron trained using backpropagation for classification is proved to be a minimum mean squared-error approximation to the Bayes optimal discriminant functions. A simple technique for sensor fusion is shown to provide a statistically significant improvement in performance using absolute range and forward looking infrared imagery for target detection over the single sensor case. Keywords: Theses, Neural nets, Pattern recognition, Target detection, Multisensors, Statistical decision theory, Artificial intelligence. (RH)