基于形态学的两种高光谱目标探测改进算法

Two modified target detection algorithms based on morphology for hyperspectral imagery

基于高光谱图像的目标探测主要利用高光谱图像具有连续光谱信息的优势,通过目标与背景的光谱信息差异进行探测,空间形态信息则很少使用。然而在一些光谱分析算法的基础上充分考虑图像像元之间的空间关系能够克服这些算法存在的不足。本文针对约束能量最小化(CEM)算法采用全图信息估计背景信息、正交子空间投影(OSP)算法难以准确构造背景子空间的不足,引入目标的空间尺寸信息,然后通过形态学开运算滤掉可能存在的目标以此构造一个精确的背景,进而在统计精确背景信息的基础上使用OSP、CEM两种算法进行目标探测。最后本文采用三维受试者操作特性(3D ROC)曲线进行探测精度的评价,克服了传统2D ROC曲线评价探测精度的不足。实验结果表明,基于形态学的CEM和OSP算法有效地降低了虚警率,提升了探测效率。

Hyperspectral image-based target detection makes good use of the advantage of continuous spectral information of hyperspectral image,distinguishing the target and background mainly by spectral difference.However,fully considering the spatial relationship between image pixels in some spectra anal-ysis algorithms can overcome the shortcomings of these algorithms,such as the lack of constrained energy minimization （CEM）algorithm estimating the background information with the information of full map and orthogonal subspace projection （OSP）algorithm being difficult to accurately construct background subspace.In this paper,spatial dimension information of targets are introduced and possible targets are filtered out by means of morphological opening operation to construct an accurate background,based on which OSP and CEM algorithms are implemented respectively for target detection.The 3D ROC curve is used to evaluate the accuracy of the detection results to overcome the shortcomings of traditional 2D ROC curve.The experimental results show that morphology-based CEM and OSP algorithms can effectively re-duce the false alarm rate and improve detection efficiency.