Anomaly Detection Methods For Hyperspectral Image Based On Variational Autoencoder

Hyperspectral remote sensing technology is capable of collecting images with extremely high spectral resolution,which enables this technology to distinguish ground objects with small spectral differences,so it has become a trending topic in the field of remote sensing.In the application research of hyperspectral remote sensing,hyperspectral anomaly detection can find pixels that are significantly different from the background without the prior information of the target,thus becoming the research direction with the most practical application value in hyperspectral image analysis tasks.At present,hyperspectral anomaly detection has been used in many fields such as mineral exploration,military camouflage detection,disaster rescue,etc.Therefore,it is necessary to conduct in-depth research on anomaly detection.However,when accomplishing the detection tasks,the hyperspectral anomaly detection is also confronted with such challenges as high data dimensions,high data redundancy,and uneven proportions of abnormal pixels.So far,many researchers have proposed different hyperspectral anomaly detection methods.These methods contribute to solving these problems to some extent,while many current methods only obtain detection results through a large number of analyses of disordered spectral dimension data,ignoring the important characteristics of spatial neighborhood information.Therefore,in order to effectively utilize the spatial and spectral dimension information in hyperspectral images,two hyperspectral anomaly detection methods are designed on the basis of variational autoencoders,and then an anomaly detection system is built on these two methods.The main points of this thesis are as follows:1.This thesis designs a hyperspectral anomaly detection method based on 3D convolutional variational autoencoder.In order to effectively describe the low-dimensional manifold structure of the background pixels of the hyperspectral image,this thesis uses a 3D convolutional variational autoencoder to represent the hyperspectral image,and at the same time introduces an adaptive weight regular term to increase the difference between the background and anomalous pixels.Finally,the RX detector is used to detect the anomaly of the residual between the background of the hyperspectral image and the original input,and then reach the final detection result.Through the experiments on real hyperspectral image data sets,the results prove that this method is capable of detecting abnormal targets in hyperspectral images with good detection effects.2.In order to further utilize the spatial information in hyperspectral images,an anomaly detection method for hyperspectral images is presented based on graph variational autoencoders.Firstly,this method uses the super pixel segmentation algorithm to segment the hyperspectral image and construct an adjacency matrix,so as to evaluate the similarity between pixels.Secondly,a variational autoencoder is used to reconstruct the spectral vector of the hyperspectral image,and meanwhile the spatial similarity of the image is shared in the feature space through the graph regularization term.Finally,the reconstructed background and the original input are used to obtain the spectral error map,and then the attribute filtering is used to further refine the detection results.Performed on four sets of abnormal target data with different shapes and different background complexity,the experiments show that the method has good anomaly detection performance.3.In order to facilitate users to detect anomaly targets in hyperspectral images,a goodlooking and easy-to-use hyperspectral image processing and anomaly detection system is built on.Net Framework by using C# language,realizing such functions as image visualization,data preprocessing,anomaly detection,etc.In addition,the practicability of the system and the effectiveness of the methods in this thesis are verified by an application example of anomaly detection in hyperspectral images.