Anomaly detection based on an iterative local statistics approach

Abstract

One of the most challenging problems in automatic target detection is associated with the large variability of background clutter and object appearance. In this paper, we propose an anomaly detection approach which does not rely on an exhaustive statistical model of the targets, but rather on the local statistics of the data and possibly on some a priori information regarding the sizes and shapes of targets. Iterative procedures of feature extraction and anomaly detection are carried out, gradually reducing the false alarm rate while maintaining a high probability of detection. The background is characterized in a feature space of principal components, and a single hypothesis scheme is used for the detection of anomalous pixels. Morphological operators are subsequently employed for extracting the sizes and shapes of anomalous clusters in the image domain, and identifying potential targets. The robustness of the proposed approach is demonstrated with application to sea-mine detection in sonar imagery.