Research On Method For Target Detection In Hyperspectral Imagery

Hyperspectral imagery(HSI),a 3D data cube((with spatial-spatial-spectral components),provides a wealth of spectral information to uniquely identify various materials by their reflective spectrum,which makes target detection in HSI has drawn lots of attention due to its advantage over the last two decades.Anomaly detection in HSI can distinguish targets of interest from the background,requiring no prior spectral information of targets,which is becoming more promising.Thus,researches on the anomaly detection methods are of great importance to detect unknown targets more efficiently.After projection transformation in spectral dimension,pixels with abnormal spectrum are picked out as targets in most existing anomaly detection methods,while information concealed in space dimension.is seldom taken into consideration.Aiming at improve the performance of detection algorithm by combining spectral information with special information for target detection,some efforts have been carried out in this paper.The main works are listed as follows:(1)A method based on local linear fitting for anomaly detection is proposed.Since the background-pixels are much more than target-pixels,pixels in the cluster with a certain amount samples are labeled as background-pixel,while anomaly pixels can not be labeled in this method.Then,the residual of locally linear fitting is taken as the measurement of anomaly degree.The unlabeled pixels with relative large residual are detected as the targets.Without assuming the distribution model of the background pixels,the proposed method improve the performance of target detection,as the backgroundpixels do not always suit with the assumed model in other methods.(2)An anomaly detection approach based on background segment is proposed.The approach can be considered as a two-step detector.According to the spectral similarity,the scene is divided into several connected regions(background region),which is achieved by spectral region growing in the first step.In the second step,the neighbors of a pixel are determined by the topological relationship between the pixel and the background region.Whether the pixel is a target-pixel depends on its residual when it is linearly fitted by its neighbors.Experiments show that the proposed algorithm can detect the target at a low false alarm rate.(3)A detection scheme for ship target is designed.According to the different reflection characteristics in near infrared bands,sea area is extracted from the scene,which is the target area for ship detection.Then RX detector is used to get the anomaly pixels.Covered by anomaly pixels,the objects with a certain geometry shape are detected as ship targets.Besides,the detector based on background segment is also used for ship detection.Compared with the former scheme,the process of detector based on background segment is more simple.