CONTEXT-DRIVEN LANDMINE DETECTION USING SEMI-SUPERVISED MULTI-TASK LEARNING

The proliferation of landmines continues to be a problem of worldwide humanitarian urgency. While airborne sensors have demonstrated significant utility in covering a wide area at a high stand-off distance, the variety of deployment methods, environments, mine types, and operating conditions continue to pose challenges in the context of landmine detection requirements. In this effort, a context-driven approach to the landmine detection architecture is proposed with a focus on transitioning new mathematical and statistical tools for applied image processing that are highly relevant to addressing landmine detection challenges. We propose a framework that provides: texture-based context for anomaly detection; false alarm reduction through semi-supervised learning and multi-task learning where data associations are learned in the feature space and the classifier parameter space; mine field association through graph-based diffusion; and a mechanism to explicitly incorporate analyst feedback to optimize performance under new operating conditions. Computational efficiency of the underlying methods will aid in pursuing real-time implementation, test and evaluation in Phase II.