Self-paced relaxed collaborative representation with adaptive clustering for hyperspectral anomaly detection

Hyperspectral anomaly detectors based on collaborative representations have received much attention for their effectiveness. Recently, fully utilize the band features, the relaxed collaborative representation detector (RCRD) has been proposed and its superiority has been verified. However, it does not consider the interference of noisy bands and redundant dictionary atoms in the representation process for each pixel under test (PUT). In view of the issues, this article proposes a self-paced relaxed collaborative representation detector (SP-RCRD). Unlike the predefined weights approach, the self-paced learning (SPL) strategy is used for iterative optimization, allowing so that the set of bands to be involved in the representation progressively, from “good” to “bad”, as the model is solved. In addition, a weighted version is proposed for differences between dictionary atoms. Moreover, in the absence of any prior knowledge, an adaptive clustering method is proposed to construct dictionaries, which can greatly improve the generalization and automation of our proposed detection method. Experiments show that the proposed method outperforms existing methods. • Self-paced relaxed collaborative representation is proposed for hyperspectral anomaly detection. • A weight matrix is constructed to adjust the importance of these dictionary atoms. • An adaptive clustering method is designed to construct robust background dictionary without prior knowledge.