Neurodynamic networks for recognition of radar targets
H.A. Babri · Scholarly Commons (University of Pennsylvania) · 1992
In this dissertation, advanced concepts and methods of pattern recognition are brought to bear upon the problem of recognizing radar targets in presence of distortion (i.e. identification irrespective of target aspect, range, or position), and in uncertain environments. The neurocomputational approach is attractive, particularly in solving this problem, because it is not possible to define a simple analytical expression to relate the shape of the target with its vector field response to radar illumination. Neural networks can make connections between two intricately related concepts by extracting relevant features from available examples and generalizing from them to new examples. To achieve appropriate generalization is difficult because targets of interest are usually very similar, and because practically not all possible targets can be taught to the network. Hence a network with sharp cognitive ability is required to discriminate between familiar (learned) and unknown (novel) targets as well as between the learned targets themselves. A novel composite network consisting of banks of feature-forming feedforward networks and feature binding periodic attractor networks (PANs), is proposed for the solution of this problem. The feature-forming networks do local discrimination of features by processing segments of target signatures, followed by PANs which provide the cognitive mechanism by their ability to bifurcate between different attractors depending upon the familiarity of the input. Experimental scattering data from three scale model targets of the B52, the B747 and the NASA Space Shuttle is used to test the performance of the composite network. When taught with range profiles from two targets, the network is able to recognize known (or learned) targets almost perfectly and discriminate against the third novel target as well as spurious inputs or noise with high reliability. The importance of multisensory information (polarization in this context) in achieving the optimum number of segments as well as segment size naturally falls out of this approach. This is demonstrated indirectly by using simulated multisensory data. A detailed example to design composite networks with given specifications is described. The successful demonstration of the power of neural networks computing with diverse attractors to achieve higher level cognitive functions is a major contribution of this work. The feature forming networks provide the network with robustness and generalizing ability. The periodic attractor network (PAN), which integrates the decisions of feedforward networks with cognitive prudence is reasonably robust to imperfections or perturbations in its synaptic weight matrix. That the PAN lacks robustness against elemental failure suggests that the brain may be using a different mechanism for feature binding, perhaps populations of phase locked neural oscillators.