Anomaly Detection In Hyperspectral Images Based On Tensor Decomposition And Sparse Low-rank Representatio

With the progress of remote sensing science and technology,the spectral resolution and spatial resolution of hyperspectral images have been greatly improved and hyperspectial image is more and more informative.Anomaly detection of hyperspectral image is used to detect pixels with significant differences from the spectral information of surrounding background pixels,which has very important significance in various aspects such as resources utilization,military exploration,social life and so on.At present,many effective algorithms have been proposed in the field of anomaly detection of hyperspectral images,but the traditional detection methods are usually based on matrix form,which destroys the original spatial structure of hyperspectral images to a certain extent.In the practical application scenarios,the detection results are often affected by factors such as noise,and the stability of anomaly detection model cannot be guaranteed.Therefore,to design an efficient,accurate and robust hyperspectral image anomaly detection model is an important research hotspot in the field of remote sensing.This paper made a detailed analysis of the spatial-spectral structure and prior information of hyperspectral images and proposed a new method based on low-rank sparsity representation and tensor decomposition.The main research work of this paper including 3 aspects.1.We proposed an anomaly detection method based on the linear total variation and low-rank representation with tensor truncated nuclear norm.In our method,the original hyperspectral image is divided into a low-rank background tensor and a sparse abnormal tensor.We use tensor truncated nuclear norm to constraint the background tensor and tensor tube-sparse regularization to constraint the anomaly tensor.Meanwhile,as edge smoothing prior is inherent to the image,the linear total variation regularization is used to further optimize the model based on this prior information.The results of anomaly detection experiments on three hyperspectral datasets show that our method can get a good performance of hyperspectral image anomaly detection.2.A hyperspectral image anomaly detection model based on tensor low-rank total variation under mixed noise is proposed.In the imaging process of a hyperspectral image,there could be some influence of noise with various types and different intensities.Takes full consideration of the inherent global spectral correlation prior and spatial local smoothness prior of the hyperspectral images,our method separates the mixed noise using l1-norm,and uses a tensor low-rank total variation regularization taking it as the low-rank constraint of overall image.Similarly,we detected the anomaly target accurately using tensor tube-sparse regularization.The experiment results show that our method can obtain accurate anomaly detection result under the influence of the mixed noise.3.We developed an interactive hyperspectral image anomaly detection system.The system contains three functional modules: file management module and two anomaly detection modules in the usual case and noised case respectively,and we detailed introduce the framework of the system in this paper.Our system is easy to use with friendly user interface,interactive processing and interpretation functions.