Adaptive window based collaborative representation for hyperspectral anomaly detection with fusion of local and global information

Hyperspectral anomaly detection using collaborative representation (CR) has attracted high interest in recent years. Ignoring global information and the use of fixed dual window, which is inappropriate for targets with different sizes, are some disadvantages of the existing methods. In this paper, the adaptive window based CR, called as AWCR, is proposed, which utilizes the results of two segmentation maps with different numbers of superpixels to find appropriate size of inner and outer windows for each test pixel. In addition to local information contained in adaptive dual windows, two individual dictionaries are obtained for background and anomaly subspaces from the whole image to provide the global information. Both local and global residual terms are fused to result in the final residual term in AWCR. The experiments show high detection performance with a reasonable computation time for AWCR compared to several serious competitors.