层次优化目标与背景的匹配滤波高光谱目标探测研究

匹配滤波(MF)算法通过比较待探测像元到目标类和背景类中心的马氏距离来判断目标和背景,但是无法知道高光谱图像背景和目标的相关统计信息,只能做一些假设,导致探测结果不准确。本文在MF探测完成后,设定阈值进行分割,对背景和目标进行重新划分,获取背景和目标的相关统计信息。为了更好地获得两类的统计信息,我们利用迭代的方法逐层次获得两类的统计信息,最后利用MF算法来进行目标探测。在运算过程中为了扩大目标和背景的的差异,增加目标与背景的可分性,向像素中添加零向量来调整像素值。利用两组真实高光谱数据来进行目标探测实验,本文提出的算法与MF算法和一些传统目标探测算法相比,ROC曲线和曲线下方面积AUC的值都得到提高,能够更好地抑制背景和突出目标,表明本文提出的算法能够提高高光谱目标探测的效果和精度。

Research on detection of hierarchical optimization targetand background matched filter hyperspectral target

The matched filter (MF)algorithm can judge the target and background by comparing the Marmanian distance between the pixels to be detected and the center of the target class and the background class,but it cannot know the relevant statistical information of the background and target of hyperspectral image,so it can only make some assumptions,leading to inaccurate detection results.In this paper,after MF detection is completed,threshold is set for segmentation,background and target are reclassified,and relevant statistical information of background and target is obtained.In order to better obtain the two types of statistics,we use the iterative method to obtain the two types of statistics level by level,and finally use the MF algorithm to carry out target detection.In order to enlarge the difference between the target and the background and increase the separability of the target and the background,zero vector is added to the pixel to adjust the pixel value.Two groups of real hyperspectral data were used to conduct target detection experiments.By comparing the value of ROC curve and AUC of the area under the curve,it was found that the algorithm proposed in this paper could improve the accuracy of hyperspectral target detection compared with MF algorithm and some traditional target detection algorithms.