Real-time kernel collaborative representation-based anomaly detection for hyperspectral imagery

The kernel collaborative representation detector (KCRD) has desirable detection accuracy in hyperspectral anomaly detection. Accordingly, we propose a real-time version based on KCRD (RT-KCRD), aiming to greatly improve its detection efficiency and accuracy. As for the first target, some endeavors are made: (1) the recursion between the kernel covariance matrices at the last moment and the current moment is implemented to avoid the repeated calculation; (2) Cholesky decomposition is utilized to efficiently calculate the inversion of regularized kernel covariance matrix with symmetry and positive definition; (3) the detection values are directly imposed on the regularized matrix instead of the kernel operations, furtherly reducing the overall computationally complexity. As for the second target, it can be realized by the above operation on the regularized matrix, which represents the relationship between pixels more rationally. Experimental results on three hyperspectral data demonstrate that the proposed RT-KCRD may outperform the other detectors in hyperspectral real-time anomaly detection.