A shaped collaborative representation-based detector for hyperspectral anomaly detection

A modified version of the collaborative representation-based detector (CRD) is proposed for hyperspectral anomaly detection. In contrast to the conventional CRD, which uses a rectangular dual window, the shaped CRD (SCRD) selects the most appropriate neighbours from the dual window and discards the inappropriate ones. To this end, similarity of the neighbouring pixels to the centre is computed based on the cosine distance to utilize the local information. In addition, the low/high occurrence probability of anomalies/background exhibited in the histogram of the whole image is utilized as global information to find the closest neighbours to the background. The shaped dual window is used for linear approximation of pixels through the collaborative representation. SCRD improves the anomaly detection results with respect to some related works. Experiments on two hyperspectral images show that SCRD results in more accurate detection maps with a bit higher running time compared to CRD.