Research On Hyperspectral Anomaly Detection Technology Based On Fusion Of Spatial And Spectral Information

Hyperspectral anomaly detection has been applied to various fields as a new spectroscopic technique for observing ground objects and detecting targets,and has promoted the development of technology in related fields.However,there are often problems in hyperspectral anomaly detection,such as target misses,false targets,noise interference and inconspicuous background separation.The reason is that the large amount of hyperspectral data,high information redundancy,high noise interference and insufficient utilization of spatial information.Hence,to solve the above problems,anomaly targets and background are distinguished by one-dimensional spectral information and two-dimensional spatial information.Hyperspectral anomaly detection has become an important branch of image processing field.In this dissertation,for the problem of underutilization of spatial information by hyperspectral anomaly detection techniques,four hyperspectral anomaly detection methods were proposed.First,a local sliding change double window model was established.Then,a hyperspectral anomaly detection method with local joint subspace processing and support vector machine was proposed by combining the joint subspace processing and support vector machine classification methods.Second,the traditional visual attention model was introduced.Then,a novel method based on visual attention and background subtraction with adaptive weight was proposed.Third,the hyperspectral anomaly detection method based on harmonic analysis and low-rank decomposition was proposed by combining harmonic analysis theory.Finally,to improve the accuracy of low-rank and sparse matrix decomposition,a hyperspectral anomaly detection method with spectral-spatial complementary decision fusion was proposed.In this method,a three-dimensional Hessian matrix transform,low-rank and sparse matrix decomposition based on truncated kernel norm and spatial attribute filtering were introduced.The above methods effectively improve the performance of hyperspectral anomaly detection.The specific research of this dissertation is as follows.To solve the problems of inaccurate evaluation of background statistical features and contamination of background statistical features by anomaly pixels in hyperspectral anomaly detection,a hyperspectral anomaly detection method based on local joint subspace processing and support vector machines was proposed.First,a local sliding change dual window model was designed based on the traditional dual window model to incorporate spectral-spatial information.In this model,the mahalanobis distance detection and the spectral angular distance detection of subspace projection were combined to obtain the local subspace score.Then,to further improve the detection probability,the support vector score was obtained by binary classification of the dimensionality-reduced hyperspectral images with a trained support vector machine model.Finally,the anomaly detection result was obtained by fusing local subspace score and support vector score.To solve the interference of noise and the underutilization of salient features of hyperspectral images,a hyperspectral anomaly detection method based on visual attention and background subtraction with adaptive weight was proposed.First,a hyperspectral visual attention model was constructed to extract the saliency feature map of the reduced-dimensional hyperspectral image.Second,the background information in the salient feature map was removed by the curvature filtering to obtain the initial anomaly region map.Finally,to utilize the spectral information of the hyperspectral image,an adaptive weight was applied to the initial anomaly region map to obtain anomaly detection result.To solve the problem of the influence of redundant information and isolated noise in hyperspectral images on the detection results,a hyperspectral anomaly detection method based on harmonic analysis and low-rank decomposition was proposed.First,to reduce the dimensionality of hyperspectral images and remove redundant information,a single-pixelbased harmonic analysis was adopted to extract low-order harmonic images of the hyperspectral image.Second,a new dictionary construction method was proposed to construct a background dictionary by guided filtering and difference operator.At the same time,the initial smoothed image with isolated noise suppressed was obtained.Then,to use the low-rank properties of the background and the sparse properties of the anomalous target in the hyperspectral image,an initial smoothed image was used instead of the original hyperspectral image for low-rank decomposition to obtain the sparse matrix.Finally,the anomaly target was extracted from the sparse matrix.To solve the problem of insufficient complementary spectral and spatial dimensions,a hyperspectral anomaly detection method based on the fusion of spectral-spatial complementary decisions was proposed.In the spectral dimension,first,the directional feature map was obtained by a 3D Hessian matrix transformation to smooth the background features.Second,to accurately separate sparse matrix containing anomaly targets from the directional feature map,low-rank and sparse matrix decomposition based on truncated kernel norm were used to obtain the sparse matrix.Then,the initial detection map was obtained from the sparse matrix by the mahalanobis distance.In the spatial dimension,attribute filtering was used to suppress the background and extract the spatial features of the hyperspectral image.Then,the spatial feature images were fused to obtain a spatial weight map.Finally,to take advantage of the spectral dimension and spatial dimension,anomaly detection result was obtained by fusing the initial detection map and the spatial weight map.In the experiments,synthetic and real-world datasets were used.The above proposed methods were compared with other state–of–the–art methods qualitatively and quantitatively.The experimental results demonstrated that the proposed methods were effective and superior hyperspectral anomaly detection methods.