LOW-RANK AND SPARSE MATRIX DECOMPOSITION WITH ORTHOGONAL COMPLEMENT SUBSPACE PROJECTION FOR HYPERSPECTRAL ANOMALY DETECTION

摘要

Anomaly detection is an active field in the hyperspectral imagery (HSI) processing, which distinguishes anomaly targets from complex backgrounds without a priori information. Recently, the low-rank and sparse matrix decomposition (LRaSMD) technique, which assumes the backgrounds are low-rank and the anomalies are sparse, was employed in hyperspectral anomaly detection and achieved competitive performance. In practice, some sporadic background pixels are divided into the sparse part. Thus, the LRaSMD method could encounter false alarms because it merely depends on the sparse term. To alleviate this problem, we propose a matrix decomposition-based orthogonal complement subspace projection algorithm (MDOCSP). Specifically, the preliminary sparse and low-rank parts of HSI were extracted by LRaSMD. Then a band-wise ratio was implemented between the anomaly and background to suppress the sporadic background pixels. Moreover, we projected the enhanced anomaly onto the orthogonal complement subspace of background to further reduce the false alarms. The proposed method was compared with five detectors, including Reed-Xiaoli (RX) detector, subspace RX (SSRX), cluster-based anomaly detector (CBAD), collaborative representation detector (CRD), LRaSMD and LRaSMD-based Mahalanobis distance (LSMAD) on a benchmark dataset. Corresponding to each competitor, it has the detection performance improvement of 10.66%, 21.29%, 3.9%, 3.41%, 0.49%, and 6.19%. respectively. Experiments demonstrated that MDOCSP outperforms several state-of-the-art detectors.