Anomaly detection for hyperspectral images based on improved RX algorithm

An improved local RX anomaly detection algorithm is proposed. It firstly projects the images onto the background orthogonal subspace to make local data closer to multivariate normal distribution. Then for every tested pixel in the center of the sliding local window, the bands used in RX detector are chosen adaptively. To avoid the influence of anomaly information on the background characteristic statistic, the anomalous pixels in the local background are removed and the covariance matrix is calculated using real background pixels. Finally the RX detector is used to calculate the anomalous degree of every tested pixel. Experimental results indicate it is robust and has good anomaly detection performances under complex unknown background.