Estimation of dynamic detector confidence thresholds in SAS imagery using mixture models

Abstract

As machine learning has matured over the years, more and more safety critical tasks have been entrusted to computers. Automated target recognition (ATR), the problem of identifying explosive hazards on the seafloor, is one such task that is moving toward human-supervised, and eventually, completely human-free methods. The ATR problem inherently suffers from extreme class imbalance, in that the likelihood of finding an explosive hazard in a given region of seafloor is small, but the penalty for overlooking such an explosive hazard is life threatening. In this study, we develop and test an unsupervised suppression methodology based on the confidences obtained from an existing prescreening algorithm, the combined RX detector. By treating the unlabeled set of prescreener confidences as a mixture of Laplace-distributed random variables, we can estimate the parameters of the mixture components and use this information to compute a robust threshold in prescreener confidence, below which it is safe to discard the associated ROI as benign.