Anomaly Detection and Background Suppression

Abstract

Anomaly detection is generally carried out in a blind environment with no provided a priori knowledge. With the unavailability of ground truth, a general approach is performed by visual inspection, which is probably the only means of evaluating its effectiveness, in which case background information provides an important piece of information to help image analysts interpret the results of anomaly detection. In addition to addressing issues of anomaly characterization discussed in Chap. 14 and discrimination and categorization in Chap. 15, it is also interesting to note that the background suppression issue has never been explored in anomaly detection. Despite a brief study on background suppression of anomaly detection via synthetic image experiments being presented in Sect. 5.4 to demonstrate how crucial background suppression is for anomaly detection, a full analysis on background suppression for anomaly detection seems missing. This chapter extends this study via real world applications to demonstrate how to evaluate anomaly detection via various degrees of background suppression. It decomposes anomaly detection into a two-stage process where the first stage is background suppression to improve anomaly contrast against background, and is then followed by a matched filter to increase anomaly detectability by enhancing intensities of anomalies. To see background suppression progressively changing with data sample vectors, causal anomaly detection is further developed to see how an anomaly detector performs on background suppression sample by sample via spectral correlation varying with sample vectors. Finally, a 3D ROC analysis is used to evaluate the effect of background suppression on anomaly detection.