高光谱图像中基于端元提取的小目标检测算法

针对高光谱图像中小目标检测问题,提出了一种基于端元提取的目标检测算法。该算法利用主成分分析的变换矩阵来构造投影算子,把原始图像投影到该算子构成的正交子空间后,大概率的背景信息得到抑制,从而突出了小概率的目标;在完成背景信息抑制的基础上,利用迭代误差分析方法进行端元的自动提取;根据所提取出的目标端元的光谱,结合光谱角度匹配技术完成目标物的检测。为了验证新方法的有效性,利用高光谱数据进行了实验研究,并与经典的RX算法的检测结果相比较。实验结果表明提出的基于端元提取的算法不需要目标的任何先验知识就能达到比较好的目标探测效果,对RX算法检测效果不太理想的小目标也能准确识别。

A Small Target Detection Approach Based on Endmember Extraction in Hyperspectral Image

A new target detection algorithm in hyperspectral imagery based on endmember extraction method is introduced.The purpose of our target detection is to locate and search for targets which are relatively small with low probabilities in an image scene.We use the transformation matrix of principal components analysis to construct an orthogonal subspace projection operator that projects the hyperspectral image onto a subspace,which is perpendicular to the space spanned by transformation matrix.In this subspace,background information is effectively suppressed and small targets become obvious.So,endmember spectra of targets can be extracted using iterative error analysis method.Then,we segment the targets according to the spectral angle between selected endmember spectra and each pixel vector of hyperspectral image.The proposed algorithm was studied using real hyperspectral data and compared with RX algorithm. Experimental results show that the algorithm can effectively and reliably detect the small target without prior knowledge.