Detection of subpixel anomalies in multispectral infrared imagery using an adaptive Bayesian classifier

The detection of subpixel targets with unknown spectral signatures and cluttered backgrounds in multispectral imagery is a topic of great interest for remote surveillance applications. Because no knowledge of the target is assumed, the only way to accomplish such a detection is through a search for anomalous pixels. Two approaches to this problem are examined in this paper. The first is to separate the image into a number of statistical clusters by using an extension of the well-known k-means algorithm. Each bin of resultant residual vectors is then decorrelated, and the results are thresholded to provide detection. The second approach requires the formation of a probabilistic background model by using an adaptive Bayesian classification algorithm. This allows the calculation of a probability for each pixel, with respect to the model. These probabilities are then thresholded to provide detection. Both algorithms are shown to provide significant improvement over current filtering techniques for anomaly detection in experiments using multispectral IR imagery with both simulated and actual subpixel targets