Canonical analysis for increased classification speed and channel selection

Abstract

The quadratic form can be expressed as a monotonically increasing sum of squares when the inverse covariance matrix is represented in canonical form. This formulation has the advantage that, in testing a particular class hypothesis, computations can be discontinued when the partial sum exceeds the smallest value obtained for other classes already tested. A method for channel selection is presented which arranges the original input measurements in that order which minimizes the expected number of computations. The classification algorithm was tested on data from LARS Flight Line C1 and found to reduce the sum-of-products operations by a factor of 6.7 in comparison with the conventional approach. In effect, the accuracy of a twelve-channel classification was achieved using only that CPU time required for a conventional four-channel classification.