A generalized regression neural network (GRNN) scheme for robust estimation of target orientation using back-scattered data

The ill-conditioned nature of electromagnetic inverse scattering problems calls for newer approaches to the old problem. The problem of estimating the orientation of a conducting target is successfully handled by a generalized regression neural network (GRNN). The training data set consists of backscattered RCS data measured when the target is exposed to an incident single frequency TM plane wave. Noisy data sets are then provided to increase system robustness. The network response proves to be highly robust to problem setup non-idealities such as target scaling, sensor misplacements, sensor noise and template deformations. Further investigations show that even reducing the total number of sensors does not affect the general capabilities of the network much but slightly reduces the robustness, especially when the target is deformed. Also, confining the sensors to one half plane still maintains the generalization capability of the network but reduces its robustness. Time domain schemes, such as range profiling techniques, can utilize of this method to overcome their difficulties in estimating the orientation of the target.