Anomaly Target Detection In Hyperspectral Image Based On Background Purification Via Density Peak

Hyperspectral images are different from other ordinary optical images,and have very abundant spectral information.It has attracted more and more attention both from domestic and foreign researchers.Among them,the hyperspectral anomaly detection technology which doesn’t need prior information of the target is paid increasing attention by researchers.Researchers at home and abroad have invested a lot of energy in hyperspectral anomaly detection and have achieved many results.However,the characteristics of hyperspectral data have become an important factor limiting the effect of anomaly detection.For example,the problem of information redundancy in hyperspectral data,and the problem that the same objects can have different spectral information and different objects can have the same spectral information,caused by light and atmospheric conditions,will affect the accuracy of anomaly detection.Besides,some commonly used anomaly detection methods are based on the premise that noise conform to a simple gaussian distribution.However,in fact,the noise of real hyperspectral data is a very complex structure,which is one of the reasons that affect the accuracy of hyperspectral anomaly detection as well.In practical applications,there are often two situations,that is,the prior information of some pixels of an image is easy to obtain,but the prior information of most other pixels is not,or the prior information of all pixels of an image is not easy to get.In response to these problems,the abnormal target detection in hyperspectral images is studied in detail in these two cases.The main contents of this paper are as follows:In the case that part of the prior information is known,this paper proposes to use the Extreme Learning Machine to extract features of hyperspectral data in response to the problem of information redundancy caused by the excessively high spectral dimension of hyperspectral data and the phenomenon that the same objects can have different spectral information and different objects can have the same spectral information.This method converts high-dimensional hyperspectral data into low-dimensional features.Meanwhile,it retains the feature information of hyperspectral data points to the maximum.Although general unsupervised algorithms and recently popular deep learning networks can extract the deep features of data,because these methods require a long time,it can shorten the detection time and improve the detection accuracy by using the extreme learning machine of single hidden layer,which has strong learning ability and fast convergence speed.Moreover,noise and anomalies are also factors that affect the accuracy of anomaly detection.In order to solve this problem,a pure background extraction algorithm based on density peaks is proposed to calculate the local density of pixels in dual windows,which not only avoids the influence of noise and anomalies on the detection results,but also makes full use of the spatial and spectral information of hyperspectral data.Finally,the anomaly detection algorithm based on collaborative representation is used for detection,which avoids making assumptions about the background and is more in line with real hyperspectral data.By detecting three real hyperspectral data and comparing them with five other classic anomaly algorithms,this paper verifies that this method can improve the detection accuracy of data sets acquired by different sensors to a certain extent.When the prior information of all pixels is unknown,the traditional anomaly detection method based on low-rank sparse decomposition only considers the problem of the simple gaussian distribution of noise.In this paper,a mixed gaussian noise model is introduced into low-rank sparse decomposition.Due to the influence of light and atmospheric conditions,the noise in hyperspectral data is not only a simple gaussian distribution,but a more complex mixed noise.Considering the mixed gaussian noise,it is more in line with the real hyperspectral data,which can improve the accuracy of anomaly detection.Secondly,traditional algorithms based on anomaly detection only consider sparse components and ignore the low-rank components.In fact,low-rank background components also contain some abnormal information,so further detection of low-rank background components is also beneficial to improve the detection accuracy.In addition,there may be some noise in the low-rank background components,so in this paper,an anomaly detection algorithm based on density peak background purification is used to detect the low-rank background components,avoiding the influence of noise.At last,an adaptive fusion method based on entropy is proposed in this paper,which combines the detection results of low-rank and sparse components to avoid the influence of the weight set by experience on the detection results.