Independent Classifiers in Ontogenic Neural Networks for ATR

This paper investigates independence of classifiers in neural networks for high range resolution radar target recognition. Independent classifiers are used to select distinguishing features for synthesis of ontogenic neural networks (networks that generate their own topology during training). A class of nonorthogonal classifiers is defined and their classification properties are investigated. Radial basis functions and wavelet transforms are used to preprocess the radar signal data. Data preprocessing is used to minimize the effect of noise, phase shift, and scale change of the radar signal. Simulation results based on synthetically generated aircraft radar images showed promise for automatic target recognition. Introduction Automatic target recognition requires an accurate recognition algorithm which can be implemented on real time hardware. The problem of pattern recognition can be efficiently solved on a dedicated neural network (NN), trained off line, using an extensive data base. Neural networks use many paradigms and data organizations which are potentially useful for target recognition. Radial basis functions (RBF) were shown to provide an arbitrarily close approximation to any continuous function [Girosi]. In addition, [Chen] demonstrated that any continuous function can be used as an activation function in RBF neural networks (RBFN), as long as it is not an even polynomial. Recently [Zhao] used RBFN for automatic target recognition based on Kuband radar range profiles of three military aircraft. He demonstrated that by using a Fourier transform and non-coherent amplitude averaging, stable and shift invariant features were obtained which significantly improved the classification quality. In addition, classification based on RBFN was more accurate than using minimum-error Bayesian classifiers. In this paper, we use RBF to produce a minimum set of independent classifiers in the input transformation space. The input transformation space includes the original signal, its amplitude range, average value, standard deviation, and additive information cost functions (Shannon entropy, log energy, and l norms.). In addition a Fourier transform and a Haar wavelet transform of the original signal were used to enhance the input transformation space. In this space, a selection of features for nonorthogonal classifiers were performed. The classification results were compared with classification based on the orthogonal RBF classifiers obtained in the same input transformation space. NN training and test were performed using model based, synthetically generated, high range resolution radar aircraft images.