Identifying anomalous objects in SAS imagery using uncertainty

Object detection in modalities such as synthetic aperture sonar (SAS) is affected by the difficulty of acquiring a large number of training samples. If object classes not present in the training dataset are detected during testing, they can be mis-classified as one of the training classes. This increases overall false alarm rate and affects operator reliability and trust in the detection algorithm. Previous work showed that classification algorithms are often overconfident in their predictions and hence cannot reliably flag image regions about which the algorithm is uncertain or which need further sampling or processing. This paper describes object detectors based on SVMs and Gaussian Processes for SAS imagery, followed by probabilistic calibration of detector confidence scores. The entropy or uncertainty of these scores is then used to identify low-confidence regions and indicate the presence of previously unseen or anomalous objects.