Research On Hyperspectral Anomaly Detection Algorithm

Hyperspectral remote sensing technology is a non-contact detection technology since the 1960s.Usually,its spectral resolution range is from 10nm to 20nm.There are many bands in hyperspectral image,which can provide a lot of information for target detection,recognition,classification and other operations.Hyperspectral anomaly detection belongs to the target detection of unknown prior information,which is more practical than the target detection of known prior information,so it has attracted widespread attention from researchers.Based on the theory of characteristics of HSI data,sparse representation and collaborative representation,this paper focuses on the model construction and purification of background dictionary.Meanwhile,combining with the background dictionary,two HSI anomaly detection algorithms based on sparse representation and collaborative representation are proposed.Simulation results show that the proposed algorithms effectively improve the accuracy of anomaly detection.Aiming at the low precision of HSI anomaly detection,an anomaly detection algorithm based on dictionary random subspace collaborative representation is proposed.In order to obtain the possible types of objects in HSI,the mean shift algorithm is used for clustering and the classification results are optimized to avoid the pollution of the background dictionary.In view of the small proportion of anomalies and noises in HSI,the categories with small number of image elements after clustering are often composed of anomalies or noises.Thus,eliminating these categories is conducive to building a pure background dictionary.In order to select more similar pixels,the distance between each pixel and its cluster center is calculated in the remaining terrain categories,and the nearest pixel is selected as the atom to construct the background dictionary.Finally,the background dictionary is used to represent each pixel in the original HSI and output the final abnormal detection results.On the issues of the abnormal pixel pollution in background dictionary,a new HSI anomaly detection algorithm based on the construction of possible anomaly dictionary and sparse representation is proposed.Firstly,the possible anomalies and noises in the original HSI are removed by bilateral filtering based on spatial distance weight and spectral Euclidean distance weight.Secondly,the possible anomaly dictionary is constructed and purified by using the difference map of the original HSI and the filtered result image.Thirdly,the background dictionary is constructed and purified from the original HSI by binary K-means clustering.Fourthly,each pixel in the original HSI is sparse represented through the constructed exception dictionary and background dictionary.Finally,distinguish anomalies via using judgment criteria and output the final abnormal detection results.In order to verify the performance of the proposed algorithms,the two algorithms are simulated and compared by using the hyperspectral data set ABU and San Diego,meanwhile,using subjective and objective evaluations to realize the judgment.The results show that the two algorithms can effectively reduce the possibility of abnormal pixels in the background dictionary and improve the detection accuracy.