Blind source separation based anomaly detection in multi-spectral images

Anomaly detection from multi-spectral images is a standard image processing problem in civilian and military applications. The major difficulty of this problem lies in the complex background modeling: not only the background is usually heterogeneous (contains multiple sources, e.g. vegetation, soil, etc) but also spectrally varying even for an homogeneous background. Unlike most existing methods that rely on a Gaussian background model, we do not consider any parametric statistical model for the background; the latter is modeled by a linear mixture of multiple sources estimated using recent sparse blind source separation methods. The detection procedure is then based on contrast measures derived from the estimated mixture parameters. We numerically show that our method is more relevant than two other methods of the state of the art.