부산대학교 도서관

We introduce new methods for training an autoencoder (AE) as an unsupervised hyperspectral anomaly detector. We detail a new percentile loss (PL) that reliably constructs an accurate background model while limiting the erroneous inclusion of anomalous pixels. We also improve detection performance and reliability by introducing a cumulative detection score that incorporates statistics calculated from the ensemble of AE models generated over the history of the training process. We show improved detection performance on two data sets relative to two baseline algorithms.