Hypothesis testing for landmine detection with EMI images

The fundamental goal of any landmine detection system is to achieve a high probability of detection while at the same time maintaining a low probability of false alarm. For the problem of detection of landmines with electromagnetic induction (EMI) sensors, the performance tends to be limited by the false alarm rate, as opposed to the detection rate. In this paper, we review a statistical Bayesian approach for deriving algorithms to test the "mine" and "no mine" hypotheses which incorporates the physical nature of the response of EMI sensors as well as the statistical nature of the clutter process into the detection framework. Theoretical performance bounds are described, and the performance of such algorithms on data collected in conjunction with the DARPA Backgrounds Clutter Experiment is described.