Modified Backward Error Propagation for Tactical Target Recognition

摘要

This thesis explores a new approach to the classification of tactical targets using a new biologically-based neural network. The targets of interest were generated from doppler imagery and forward looking infrared imagery, and consisted of tanks, trucks, armored personnel carriers, jeeps and petroleum, oil, and lubricant tankers. Each target was described by feature vectors, such as normalized moment invariants. The features were generated from the imagery using a segmenting process. These feature vectors were used as the input to a neural network classifier for tactical target recognition. The neural network consisted of a multilayer perceptron architecture, employing a backward error propagation learning algorithm. The minimization technique used was an approximation to Newton's method. This second order algorithm is a generalized version of well known first order techniques, i.e., gradient of steepest descent and momentum methods. Classification using both first and second order techniques was performed, with comparisons drawn. (rh)