A Background-Purification-Based Framework for Anomaly Target Detection in Hyperspectral Imagery

Abstract

Anomaly target detection for hyperspectral imagery (HSI) is one of the most important techniques, and many anomaly detection algorithms have been developed in recent years. One of the key points in most anomaly detection algorithms is estimating and suppressing the background information. This letter proposes a background-purification-based (BPB) framework considering the role of background estimation and suppression in anomaly detection. The main idea is the acquisition of accurate background pixel set. To prove the validity of the proposed framework, the BPB Reed-Xiaoli detector (BPB-RXD), the BPB kernel Reed-Xiaoli detector (BPB-KRXD), and the BPB orthogonal subspace projection anomaly detector (BPB-OSPAD) are proposed. Both the BPB algorithms focus on accurate background information estimation to improve the performance of the detectors. The experiments implemented on two data sets demonstrate that both BPB algorithms perform better than other state-of-the-art algorithms, including RXD, KRXD, OSP, blocked adaptive computationally efficient outlier nominators (BACON), probabilistic anomaly detector (PAD), collaborative-representation-based detector (CRD), and CRD combined with principal component analysis (PCAroCRD).