Application of unsupervised neural networks to the enhancement of polarization targets in dual-polarized radar images
Andrew M. Ukrainec, S. Haykin · 2002
The authors discuss a novel approach to contrast enhancement using an unsupervised neural network with a mutual information learning criterion. The learning algorithm used is presented, which is based on minimizing the mutual information between the network outputs. The aim is to apply this neural network learning paradigm to dual-polarized images, and thus enhance an image of a radar reflector target in radar clutter. It is shown that an unsupervised neural network can be trained to provide a contrast enhancement to dual-polarized radar images that exceeds the performance capabilities of the standard principal components analysis method. It does so by minimizing a mutual information-based cost function, thus making use of the nonlinear transformations possible with a neural network.>