Research On Anomaly Detection Of Hyperspectral Image And Parallel Implementation Based On GPU

Hyperspectral image can compress the object spectral information of different wavelength into an image cube,so it has plenty of spectral information and nowadays it has been widely used on some application such as classification,target and anomaly detection.There is no need for prior spectral information in hyperspectral anomaly detection,so it can reach the goal of real-time detection because of reduce of the load for information transfer and anomaly detection has been one of the research hotspot recently.However,with all the advances in technology,the image volume is bigger and the speed of data collection become fast than the speed of data processing.Nowadays,most part of the application in anomaly detection can not satisfy the requirement of real-time processing.The processing capacity of GPU to large-scale image matrix is higher with the fast iteration of compute capability,so there are more and more application,which use GPU to accelerate computation.RX is a classic anomaly detection algorithm and it always considered as a benchmark in comparison.Nowadays,how to reach the goal of real-time detection and improve the detection performance and stability are the research focus in the task of anomaly detection.The main work contain two points in this paper.On one hand,for the problem of large amount hyperspectral data,we have realized the parallel algorithm of RX and the experiment result verify the efficient of compute density matrix on GPU.In addition,it always produce noise in the process of data acquire because the influence of atmospheric refraction and cloud layer.The derivative features and decision fusion were used to improve the detection performance and stability.On the other hand,the hyperspectral image processing software was designed and realized base on GPU,which used the basis of the completed algorithm.This paper showed the whole architecture of the software.At the same time,we have done test to most module include RX anomaly detection,CRT classification,hyperspectral image display,data format converting and stripe noise removing and so on.