This paper deals with point target detection in nonstationary backgrounds such as cloud scenes in aerial or satellite imaging. We propose an original spatial detection method based on first- and second-order modeling (i.e., mean and covariance) of local background statistics. We first show that state-of-the-art nonlocal denoising methods can be adapted with minimal effort to yield edge-preserving background mean estimates. These mean estimates lead to very efficient background suppression (BS) detection. However, we propose that BS be followed by a matched filter based on an estimate of the local spatial covariance matrix. The identification of these matrices derives from a robust classification of pixels in classes with homogeneous second-order statistics based on a Gaussian mixture model. The efficiency of the proposed approaches is demonstrated by evaluation on two cloudy sky background databases.
© 2012 Optical Society of America
Original Manuscript: May 22, 2012
Revised Manuscript: October 10, 2012
Manuscript Accepted: October 12, 2012
Published: October 31, 2012
Laure Genin, Frédéric Champagnat, and Guy Le Besnerais, "Background first- and second-order modeling for point target detection," Appl. Opt. 51, 7701-7713 (2012)