Collaborative representation with background purification and saliency weight for hyperspectral anomaly detection

Collaborative representation-based detection (CRD) has been developed in hyperspectral anomaly detection tasks and testified to be very effective; however, heterogeneous pixels in the background may affect the accuracy of linear representation and make its performance suboptimal. To address this issue, a background purification framework based on linear representation is proposed, in which an automatic outlier removal strategy based on initial coefficients is designed to purify the background. In the proposed method, the classic least squares technique is firstly adopted to obtain preliminary linear representation coefficients, which are positively correlated with its contribution to a central testing pixel. Then, using statistical analysis of the representation coefficients, purified background pixels are obtained. Furthermore, a saliency weight is applied to fully utilize the spatial information of inner window pixels. Extensive experiments with three real hyperspectral datasets show that the proposed method outperforms state-of-the-art CRD and other traditional detectors.