ANOMALY DETECTION IN HYPERSPECTRAL IMAGES VIA SUPERPIXEL SEGMENTATION AND UNSUPERVISED BACKGROUND LEARNING

We propose an anomaly detection algorithm for hyperspectral images based on Dirichlet process mixture (DPM) models. For this purpose, we first apply an unsupervised background purification procedure to data to obtain an anomaly-free background. In this procedure, we use simple linear iterative clustering (SLIC) superpixel segmentation method to remove the small regions which are assumed to be the possible anomaly candidates in the image. Then, we propose to use DPM models, namely Chinese restaurant process and truncated DPM, to estimate the background statistics and model order. In detection part, a Gaussian-mixture-model-based (GMM-based) anomaly detector is employed. We compare the detection performances of the proposed DPM methods with the variational Bayes (VB) and classical GMM in the experimental part. The detection results show that the proposed methods outperform the performance of the other methods.