Hyperspectral Anomaly Target Detection Based On Improved Sparse Representation

Hyperspectral image(HSI)contains three-dimensional information of spatial dimension and spectral dimension,which has the characteristics of high spectral resolution and integration of image and spectrum.Comprehensive spectral information and three-dimensional data structure are helpful to distinguish various objects and detect abnormal targets.The main research direction of hyperspectral abnormal target detection algorithm is to accurately detect the target pixels which are different from the background features without knowing the prior information of the target.Most of the abnormal target detection algorithms distinguish anomaly from background by building background model and using the difference between anomaly and background.How to construct a background model without abnormal target pollution is a key problem.The accuracy of background feature extraction determines the effectiveness of hyperspectral anomaly detection algorithm.The main idea of anomaly detection method based on signal sparse representation is to evaluate the recovery error of the signal by describing the dictionary of the background subspace,and to judge whether the detected pixel is the background or the target by the error.However,the algorithm based on sparse representation does not assume the statistical distribution of the data,and the background features are also obtained by random means,in which the purity of the background determines the accuracy of the detection results.Low-rank and sparse matrix decomposition(LRaSMD)divides the matrix into low-rank matrix,sparse matrix and noise matrix.The low-rank matrix corresponds to the background matrix,and the sparse matrix can be used to detect abnormal targets.We can get a relatively clean background through LRaSMD algorithm.In addition,since auto encoder has the advantages of richness,discrimination and accuracy in extracting hidden layer features,we can also use auto encoder to obtain background features.Therefore,this thesis proposes improved sparse representation algorithms for target detection,which obtains background features from two aspects: low-rank and sparse matrix decomposition and auto encoder,then builds background dictionary model from background set by sparse expression,and finally detects outliers by calculating reconstruction error.This thesis mainly includes the following aspects:(1)The basic concepts of sparse representation theory are introduced in detail.Based on the existing research results,we elaborate several commonly used sparse algorithms and make a simple comparison and analysis.The introduction of sparse representation lays a solid theoretical foundation for the improved anomaly detection algorithms proposed in the following chapters.(2)We introduce the principle of LRaSMD algorithm,which can distinguish background from anomaly,so it can greatly reduce the background pollution caused by abnormal targets.A hyperspectral anomaly detection algorithm based on LRaSMD and sparse dictionary representation is proposed.First,a relatively clean background is obtained by LRaSMD algorithm,then a background dictionary model is constructed from the background by sparse representation,and finally,an abnormal target is detected by calculating reconstruction error.(3)A hyperspectral anomaly detection algorithm based on auto encoder and sparse representation is introduced.Auto encoder fits the original input sample data through the hidden layers.Compared with traditional machine processing methods,it has stronger adaptability and does not need to select features artificially.We use auto encoding network to extract background features of hyperspectral images,and then detect abnormal targets by sparse representation.