Hyperspectral Anomaly Detection Based On Multiple Feature Low-rank Representation

Recently,the hyperspectral image anomaly detection has attracted more and more attention in the field of target detection because hyperspectral images have abundant spatial and spectral information.To improve the accuracy of hyperspectral anomaly detection,the advanced methods of low-rank representation and feature extraction are proposed in this thesis.However,the hyperspectral anomaly detection based on low-rank representation has some disadvantages.On the one hand,the spatial neighborhood information of the pixels is not considered and the spectral information is only used.On the other hand,the dictionary has an impact on the stability of the detection results.In this thesis,we study on the method of spatial-spectral information and dictionary construction,and improve the latent low-rank representation feature extraction algorithm.Meanwhile,these methods are appointment for the hyperspectral anomaly detection.In order to improve the accuracy of hyperspectral anomaly detection,we research on the spectral-spatial information,the dictionary study,and the feature extraction etc.The main achievements are as followed:(1)Based on low-rank representation,the spatial relationship between testing pixels and neighbor pixels in a single local window is constructed for hyperspectral anomaly detection.In order to obtain better spatial constraints,the multiple local window is proposed to reduce the abnormal pixels in the local window in consideration of the defects of single local window in anomaly detection.(2)Enhancd the robustness of hyperspectral anomaly detection results,the low-rank dictionary study model is proposed by adding low-rank constraints of the sparse dictionary study model.(3)The advanced model of feature extraction is proposed by introducing the Latent Low-Rank Representation to the hyperspectral anomaly detection.Meanwhile,the new transformation matrix is obtained by adding the sparse constraints to the representation coefficient.The results of anomaly detection is better than using classical detection method in feature space.