Research On Anomaly Detection Of Hyperspectral Image Based On Sparse Representation And Local Background Smoothing

Hyperspectral remote sensing is a new comprehensive discipline that has developed rapidly in the field of earth observation in a short time.With the enhancement of China's comprehensive national strength and the development of satellite applications,the application of hyperspectral remote sensing technology has good opportunities.Unsupervised abnormal target detection is one of the important applications of hyperspectral remote sensing images,because it does not require the characteristics of a priori spectral information,statistical analysis of the acquired hyperspectral remote sensing image data is used to find non-conforming backgrounds Under global or local background The abnormal target of spectral information has a wide range of applications in the fields of modern military target strike,mineral exploration,agricultural identification and environmental monitoring through the difference of spectral information between the abnormal target and the background.How to use the existing hyperspectral remote sensing data to propose a more accurate method for detecting abnormal targets that do not meet the background estimation has become a research hotspot in the field of hyperspectral image processing.Based on the lack of anomalous target prior knowledge of anomaly detection,the thesis is based on the binary classification of anomaly detection as background and target.The background statistics can be obtained to improve the performance of anomaly target detection methods.Specifically,first,in the sparse representation model,the existing algorithms usually use the original hyperspectral image data as the background training samples,which are mixed with a small number of target pixels,resulting in the low accuracy of the anomaly detection results after the learning dictionary is used for sparse representation;secondly,in the statistical model,the real hyperspectral data does not conform to the Gaussian distribution,and the distribution of target and background is greatly inaccurate Qualitatively,it causes bias and pollution to the background estimation,resulting in the low accuracy of the anomaly detection results based on the background statistics.This paper optimizes background modeling based on the two directions of sparse representation model and statistical model:(1)A sparse representation hyperspectral image anomaly detection method with two-iteration dictionary selection is proposed.The method obtains the background sample set through the low rank sparse decomposition of the matrix.In this sample set,K-SVD dictionary learning is used to complete the screening of the background dictionary,and the pure background dictionary has been used to reconstruct the hyperspectral data.Since the constructed pure background dictionary set does not contain the pollution of abnormal targets,the reconstruction error of the obtained abnonmal targets is larger,and it is easier to identify the abnormal targets.(2)A method for simultaneous detection of anomalous targets in local hyperspectral images by simultaneous similarity fusion is proposed.This method is applied to the sliding double window model,and uses different vector similarity measurement criteria to calculate the similarity of local background pixels to eliminate those pixels that differ greatly from the local background vector to avoid the pollution of local background estimation.Finally,the extracted pure background statistics are applied to the LRX algorithm to complete abnormal target detection.This paper uses ROC curves and AUC values to evaluate the performance of anomaly detection methods.Experiments are performed on three different sets of hyperspectral image data to compare different anomaly detection methods.The experimental results prove that the two methods proposed in this paper have excellent anomalies Test performance.