Research On Hyperspectral Anomaly Detection Algorithm Based On Adversarial-Learning

Anomaly detection in hyperspectral images(HSIs)is a hot topic in the field of hyperspectral remote sensing image processing,which has also been successfully applied to environmental monitoring,rare plant protection,military camouflage and hidden recognition and other fields.However,due to the influence of imaging platform error,atmospheric environment and other factors,hyperspectral images have some problems such as complex background components,data redundancy,lack of prior information.For alleviating or overcoming the negative effect of the above problems on the performance of anomaly detection,this paper deeply analyzes the data characteristics of hyperspectral remote sensing images,and designs two hyperspectral anomaly detection algorithms based on adversarial learning,respectively from the perspective of unsupervised feature extraction and semi-supervised background estimation to improve the performance of anomaly detection.The main contents and innovations of this paper are as follows:(1)A hyperspectral anomaly detection algorithm based on unsupervised feature extraction is proposed,which can effectively alleviate the interference of redundant information and complex background on anomaly detection performance.The algorithm combines the adversarial learning strategy and spectral consistency constraint to construct a feature extraction network based on encoder-decoder structure,and extracts the deep features of hyperspectral images,which can alleviate the problems of high dimension and data redundancy of HSIs.In order to improve the high false alarm rate of anomaly detection caused by complex background,the proposed algorithm utilizes attribute filtering to remove the main background information and retain the small connected area in the feature map.Then,a nonlinear mapping function is designed to locate and suppress the response of background pixels in the original HSIs with the attribute filtering result as an exponential factor,so as to enhance the discrimination between anomaly and complex background effectively.Experimental results on four real hyperspectral datasets show that the algorithm can fully express the potential features of HSIs,and significantly enhance the differences between anomaly and background.Besides,the proposed algorithm has higher detection accuracy and better background suppression ability compared with other anomaly detection algorithms.(2)A hyperspectral anomaly detection algorithm based on semi-supervised background estimation is proposed,which can to overcome the limitation of unsupervised anomaly detection algorithm due to lack of prior information,unify deep learning network and anomaly detection scheme,and achieve the balance between detection performance and training sample limitation.Aiming at the lack of prior spectral information in HSIs,this algorithm designs a scheme based on PCA and DBSCAN density clustering algorithm to construct a relatively pure background spectral sample data set by eliminating instances with high likelihoods of being anomaly.Then,the background distribution estimation network is established based on autoencoder network and adversarial learning to obtain the spectral domain feature of background spectral samples.The background representation model is generated from the background distribution estimation network.The original HSIs with anomaly and background can be reconstructed with the help of the different presentation ability of background and anomaly of the background representation model.The reconstructed data with more uniform distribution and more inclined to background distribution can be obtained by the background representation model.In order to highlight the difference between anomaly and background,the algorithm differentiates the original hyperspectral images and the reconstructed images,and then detects the anomaly on the difference data.The experimental results show that the proposed algorithm can not only break the limitation of the lack of training samples on the practicability of anomaly detection,but also has better detection performance.