Anomaly Detection of Man-Made Objects in Large Aerial Images

In this paper we present a comparison of various classifiers and features for the detection of relatively small, unknown, man-made anomalies in large, high resolution, grayscale aerial images with uniform background such as a forest. We investigate the Support Vector Machine (with and without hard negative mining), Replicator Neural Network and the Reed-Xiaoli Detector (RXD) as 1-class, unsupervised classifiers, and a number of well-known rotationally-invariant features, such as local binary patterns, local range and local mean as inputs to these classifiers. The intention is that detections made by the classifier could be used by a human image analyst to cue their attention to a small part of the large image, thereby reducing their workload. Our results indicate that the RXD classifier with the local intensity range gives the best detection rate for an acceptable false alarm rate.