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

Fully automated approaches to target detection are efficient at processing large amounts of data but often rely heavily on the training data and the model employed. Training data and modeling assumptions can be violated in the operational environments where the algorithms are applied. Alternatively, human manual target detection is accurate and adaptable due to the human ability to interpret data within its wider context. However, operational constraints and the overwhelming amount of data from modern sensors frequently preclude a fully manual approach to target detection. The SIG human-in-the-loop (HIL) active learning (AL) framework allows the operators to guide the training of the automated detection algorithm when the training and testing data statistics are mismatched. In this structured framework, the algorithm cues the operator using two specific criteria: detections that are high probability targets and detections that are highly informative for improving the classifier performance. The operator provides labels for the cued detections, and the new label information is used to retrain the classifier and improve the performance of the algorithm. At the conclusion of this Phase II effort, SIG will deliver a C/C++ implementation of the HIL/AL architecture that is ready for integration and test on the Shadow IED program.