Computation reduction of the maximum likelihood classifier using the Winograd identity

The maximum likelihood classifier is one of the most used image processing routines in remote sensing. However, most implementations have exhibited the so-called "Hughes phenomenon" and the computation cost increases quickly as the dimensionality of the feature set increases. Based on the above reasons, the recursive maximum likelihood classification strategy is more suitable for hyperspectral imaging data than the conventional nonrecursive approach. In this paper we derive some computation aspects of quadratic forms by applying the Winograd's method to three previous approaches. The new, modified approaches are approximately four times faster than the conventional nonrecursive approach and two times faster than the existing recursive algorithms.