AUTOMATED RADAR TARGET RECOGNITION BASED ON MODELS OF NEURAL NETS

This dissertation describes a new approach to target recognition, using radar returns and parallel processing based on models of neural networks. Target recognition usually consists of three steps: data acquisition, data representation (generation of feature vectors) and data classification. Two methods of target recognition are proposed and several results from their study are discussed. The methods are: (1) The use of sinogram representations as learning set in associative memory, based on models of neural nets as classifier. Such memories are known to be robust and fault tolerant, and (2) use of polarization representation for use in neural net based associative memory as a classifier.; The advantages of these methods are: (1) they represent a new approach to signal processing and target recognition, (2) they have the potential to identify targets from small fractions of the target data (robustness), (3) they are insensitive to slight degradation in system hardware (fault-tolerance), (4) they can identify targets automatically by generating identifying labels or symbols, and (5) they can be implemented efficiently using optical hardware, since optics provides the parallelism and massive interconnectivity required in neural net implementations of associative memory employing parallel processing.; Using microwave scattering data of scaled model targets, the concepts for the target recognition were demonstrated by computer simulation of a 1024 (32 {dollar}times{dollar} 32) element neural net associative memory based on the so called outer product model. The simulations show that partial input, consisting of less than 10% of the total information can identify the targets. This result illustrates the robustness of associative memory and the potential usefulness of the approach.; 2-D optical implementations of a neural net of 8 {dollar}times{dollar} 8 (= 64) binary neurons were studied. Fault tolerance and robustness are examined, using a four dimensional (8 {dollar}times{dollar} 8 {dollar}times{dollar} 8 {dollar}times{dollar} 8) clipped outer product ternary {dollar}Tsb{lcub}ijkl{rcub}{dollar} mask to establish the weighted interconnections of the net and electronic feedback based on closed loop TV system. The performance was found to be in agreement with that of computer simulation, even though aberration of lenses and the defects of the system, were present. These results confirm the practical suitability of the opto-electronic approach to the neural net implementation and pave the way for the implementation of larger networks.