Implementation of an algorithm for detection of small targets in multispectral imagery

This report describes an implementation of the RX algorithm for detection of targets in multispectral imagery. The RX algorithm assumes the background to be spatially independent Gaussian noise with a slowly varying spectral covariance matrix, which means that images must be spatially decorrelated prior to filtering. Two methods to decorrelate the background are compared: a Laplace filter and an adaptive causal AR filter with non-symmetric halfplane support set. Experimental results show that the Laplace filter leads to better detection performance for subpixel targets.