We address method of detection of anomalies in hyperspectral images that consists in performing the detection when the spectral signatures of the targets are unknown. We show that, in real hyperspectral images, use of the full spectral resolution may not be necessary for detection but that the correlation properties of spectral fluctuations have to be taken into account in the design of the detection algorithm. Anomaly detectors are useful for detecting regions of interest (ROIs), but, as they are prone to false alarms, one must analyze the ROIs obtained further to decide whether they correspond to real targets. We propose a method of exploitation of these ROIs that consists in generating a single image in which the contrast of the ROI is optimized.
© 2006 Optical Society of America
Original Manuscript: October 26, 2005
Revised Manuscript: January 13, 2006
Manuscript Accepted: January 21, 2006
François Goudail, Nicolas Roux, Ivar Baarstad, Trond Løke, Peter Kaspersen, Mehdi Alouini, and Xavier Normandin, "Some practical issues in anomaly detection and exploitation of regions of interest in hyperspectral images," Appl. Opt. 45, 5223-5236 (2006)