Rotation invariant bipolar filter design for object classification

A novel training technique for synthesizing a 3D rotation invariant bipolar filter has previously been presented (OSA Annual Meeting, October 1994). An extension of this idea is now presented. Previously, a training process using a simulated annealing algorithm for the filter was designed to yield optical correlations that were Gaussian-shaped for the preferred target and a null response for non-targeted objects. However, it is proposed that one may effectively choose other shapes for the output correlation by altering the computed filter response of the simulated annealing algorithm. A selection of basic geometric shapes may be chosen to correspond to particular classes of objects. As an illustrative example, the filter may be designed to yield a rectangular-shaped correlation in the output plane if the input object is a tank while yielding a triangular-shaped correlation for a truck input. These shaped correlations in the output plane of the optical processor may be more easily distinguished from each other and from spurious background noise through simple morphological processing of the output plane. Computer simulations of the correlation results of these filters will be presented. The design of a photorefractive-based optical correlator that implements bipolar composite filters will also be discussed.