Research On Hyperspectral Anomaly Detection Based On Sparse And Low-rank Prior Characteristics

Hyperspectral remote sensing images captures the two-dimensional spatial information about the distribution of ground objects,and at the same time provide a nearly continuous spectral curve for each pixel through hundreds of spectral bands,thus presenting the property of “image and spectrum integration”.The spectral resolution of hyperspectral images is as high as nanometers,which is sufficient to capture the diagnostic spectral characteristics of various substances,thereby providing the possibility to distinguish the fine spectral differences of ground objects.As an important application of hyperspectral image interpretation,anomaly detection aims to distinguish pixels or objects with significant spectral differences from their surrounding background,which can be regarded as a target detection problem without any prior spectral information.Hyperspectral anomaly detection technology has been successfully applied to various fields of military and national life due to its practicality.However,the detection performance of existing hyperspectral anomaly detection algorithms is usually limited by factors such as complex scene modeling,anomaly pollution in background features,waste of spatial information,spectral mixing,noise interference,and heavy computational burden.Therefore,there is an urgent need to develop some high-performance anomaly detection algorithms with high detection accuracy and low computational complexity.Although hyperspectral anomaly detection cannot provide any prior spectral information about the anomaly target of interest,there are still some prior characteristics about the background and anomaly signals hidden in the hyperspectral images,which provide the possibility for accurate anomaly detection.This paper focuses on hyperspectral anomaly detection problem,is committed to mining the sparse and low-rank prior characteristics hidden in the data,and attempts to solve the defects and shortcomings in the existing algorithms through some advanced theories and techniques,and finally achieves excellent anomaly detection performance.After introducing the research background and development status of hyperspectral anomaly detection algorithms,the main research work of this thesis is arranged as follows:(1)Focusing on the two essential shortcomings of the traditional joint sparse representation model when it is applied to hyperspectral anomaly detection,this paper proposes a robust background feature extraction method through background region segmentation-based joint sparse representation and applies it to hyperspectral anomaly detection.Through performing the eight-connected region division operation on the clustering map and then discarding small regions,this method ensures that all spectral items in each test region have similar spectral characteristics,thus overcoming the inherent shortcomings of the traditional model on the assumption of common sparsity pattern and low anomaly proportion.In addition,a novel dictionary construction technique is applied,where the dictionary atoms of one test region are composed of all spectral items in other regions with the same cluster label as this test region.By projecting the hyperspectral data into the orthogonal subspace of the background subspace spanned by the global background bases selected by the joint sparse representation model,the background component in the scene is suppressed while the anomaly signals are highlighted.Furthermore,considering the negative impact of additive noise on detection accuracy,an energy deviation-based noise suppression strategy is applied.Finally,by comparing the projection energy of each pixel on the background orthogonal subspace with the noise energy in its corresponding region,the anomaly responses of all pixel are measured.Experiments on synthetic and real hyperspectral datasets prove the effectiveness and superiority of the proposed algorithm.(2)Focusing on the problems of background incompleteness and anomaly pollution in the background dictionary used for the low-rank representation model,this paper proposes a hyperspectral anomaly detection algorithm via dictionary construction-based low-rank representation and adaptive weighting.According to the feature that the background pixels participate in hyperspectral reconstruction more frequently than anomaly pixels,by measuring the usage frequency of randomly selected dictionary atoms in each cluster for clustering reconstruction,we achieve a reasonable estimation of the background spectra in the corresponding cluster.The background spectra from all clusters are served as dictionary atoms to constitute the global background dictionary describing the entire scene,which can cover all background categories and exclude the contamination from anomaly signals.Taking into account the spectrally low-rank characteristic of the background component and the spatially sparse characteristic of the anomaly component in the scene,the global background dictionary-based low-rank representatin model effectively mines the lowestrank representation hidden in the data as the background component,and separates the sparse anomaly structure for anomaly detection.In addition,to further enhance the response difference between background pixels and anomalies,the sparse representation is perfomed on the entire data with respect to the constructed background dictionary,and the resulting representation residuals are assigned to the corresponding pixels as adaptive weight values.Reasonable background dictionary construction technology and effective adaptive weighting strategy ensure the excellent performance of the proposed algorithm to a great extent.(3)Focusing on the background interferences in the sparse component extracted by the lowrank and sparse matrix decomposition model,this paper proposes a hyperspectral anomaly detection algorithm through background suppression-based low-rank and sparse matrix decomposition.By projecting the sparse component into the orthogonal direction of the background subspace constructed from the low-rank component,this method eliminates the negative background interferences in the sparse component,reduces the false alarm rate,and finally achieves accurate anomaly detection performance.The analysis shows that compared with real anomalies,there is a significantly stronger spectral similarity between the background interferences in the sparse component and the main background signals in the scene.Accordingly,by projecting the sparse component into the orthogonal subspace of the background subspace spanned by the first few eigenvectors of the covariance matrix of the low-rank component,the background interferences in the sparse component can be effectively suppressed,while the anomaly signals are highlighted.In addition,since the lowrank component can provide a reasonable estimation of the background statistics in the scene,and results of Mahalanobis distance measurement are used to generate the adaptive weight values to further improve the separation degree between background pixels and anomalies.Experiments on three sets of hyperspectral datasets confirm that the proposed algorithm can effectively eliminate the negative background interferences in the sparse component without introducing any new parameters,and finally achieve high-performance anomaly detection.