Detection of unusual events in intermittent non-Gaussian images using multiresolution background models

A new approach to target detection filtering is developed that takes account of highly intermittent non-Gaussian backgrounds with strong phase correlations. It is shown that in some cases the departure from Gaussianity can be accounted for by the spatial intermittency of the image data and that when this is factored out by amplitude adjustment, a Gaussian process results. Spatial intermittency is modeled by considering the joint probability density of filter output (intensity) and a new measure, called local energy, which measures the background activity in the neighborhood of the filter support. A multiresolution analysis is adopted, in which multiple scales of both the filter support and the background neighborhood of local energy are considered. This approach increases the sensitivity with which targets are detected when in the vicinity of energetic regions of background. Examples of the target detection method are given for both synthetic and real imagery, showing improvements over methods that do not account for spatial intermittency.