Research On Hyperspectral Image Anomaly Detection Based On Improved Collaborative Representation

In hyperspectral image anomaly detection,collaborative-representation-based detector（CRD）algorithm realizes the detection of anomalous targets by solving the minimization problem of the l₂-norm.It does not need to assume the distribution of background and has been widely applied.However,when the pending pixel is anomalous and there is the same kind of pixel as pending pixel in the spatial neighborhood,the anomalous target becomes small and difficult to be detected.In terms of this problem,a hyperspectral image anomaly detection algorithm based on improved collaborative representation is proposed.The improved CRD algorithm collaboratively represents the pending pixel by linear combination of pixels in the spatial neighborhood,and sum-to-one constraint added to the weight of the pixel increases the stability of the solution.The Euclidean distance of pixels and the mean value in the spatial neighborhood pixels measures the anomaly degree,and the weight of each pixel is adjusted by the weighted regularization matrix of the anomaly degree.In this way,the influence of anomalous pixels in the spatial neighborhood can be reduced,so that the output of anomalous pixel increases when the above situation occurs,and the anomalous target are more easily detected.The algorithm is verified by three groups of hyperspectral image in the experiments,and the algorithm is evaluated quantitatively by the receiver operating characteristic curve,the area under curve and the detection time.Experimental results show that in the process of dual window sliding,when the same anomalous targets exist in the spatial neighborhood as the pending pixel,the proposed algorithm can achieve higher accuracy of detection than the original algorithm.In other cases,the accuracy of detection of this algorithm is basically consistent with the original algorithm.Meanwhile,anomaly detection based on improved CRD algorithm has not excess time consumption which ensures the running efficiency of the proposed algorithm.It is demonstrated that the improved CRD algorithm can effectively suppress the anomalous pixels in the spatial neighborhood and improve the accuracy of anomaly detection.