基于密度背景纯化的高光谱异常检测算法

Density background refinement-based anomaly detection algorithm for hyperspectral images

在高光谱图像异常检测中,背景存在异常像元会造成背景统计信息失真,这将导致检测结果具有较高的虚警率。针对此问题,本文提出了一种基于密度背景纯化的异常检测算法。首先计算背景中每个像元的密度;然后根据高光谱图像中背景密度远大于异常密度的特性,利用最大类间方差法将异常从背景中分离;最后,将纯化后的背景用于统计信息的估计,通过RX检测算法(Reed-Xiaoli detector,RXD)对高光谱图像进行检测。为验证算法的有效性,利用两组真实的高光谱数据进行仿真实验。实验结果表明与RXD比,所提算法在两组数据下的曲线下面积值分别提高了0.024 6和0.008 6。与当前的异常检测算法相比:所提算法有较好的接收机工作特性曲线。

In anomaly detection in hyperspectral images, a background anomaly distorts the background statistical information, leading to a high false-alarm rate. Here a density background refinement-based anomaly detector is proposed to solve this problem. The density of each background pixel was calculated first. Then, considering that the background density is much greater than the density of an anomaly in hyperspectral images, anomalies were sep-arated from background using Otsu′s method. Finally, the image was detected with a Reed-Xiaoli detector （RXD）using statistical information on the refined background. To validate the effectiveness of the proposed algorithm, ex-periments were conducted on two real hyperspectral data sets. The results show that the proposed algorithm′s area-under-curve values are 0.024 6 and 0.008 6 larger than RXD for the two data, respectively, and that it shows better receiver operating characteristics compared with existing anomaly detection algorithms.