Various linear combinations of simple matched spatial filters have been proposed in the literature to improve the discrimination in multiclass pattern recognition. It has been shown that all such approaches based on deterministic constraints can be modeled as similar matrix/vector problems, the only differences arising in the individual constraint vectors. Since the design of any of these linear combination filters (LCF) can be posed as a common matrix/vector problem, efficient iterative methods can be used to determine the LCFs. The application of one such method called the modified hyperplane (MHP) method for determining the LCF is described and its convergence behavior is numerically investigated for a set of seven patterns. It is shown that the MHP method yields correct LCFs (with rms error <0.1%) in less than ten iterations.
© 1983 Optical Society of AmericaFull Article | PDF Article