| Distortion-invariant pattern recognition is investigated in this dissertation. Correlation-based techniques, which are readily realizable through optical implementation in the form of matched-filter and joint transform correlator (JTC), are employed here. At first, a synthetic JTC is proposed, wherein, the reference image is constructed from the point spread function of an optimal filter. A recipe for the design of a composite reference for rotation-invariant recognition is furnished subsequently, that can be extended to other types of distortions. The performance of this filter-feature based composite filter is shown to be enhanced with the use of more discriminatory features extracted from the wavelet decomposition of the target image. The resulting adaptive filter is simple enough to realize and yet robust enough to cope with possible distortions. For the sake of real-time implementation of the filter, the consequences of filter discretization are considered. Discretization is addressed, in particular, using a majority-granted nonlinearity based filter formulation. |
