RX及其变种在高光谱图像中的异常检测

RX and its variants for anomaly detection in hyperspectral images

为了提高核RX算法在高光谱图像异常检测中的稳定性,将核矩阵正则化,并提出正则化的核RX算法(rkRX)。将规范化后的正则化RX算法和正则化的核RX算法融合改进,称为融合RX算法(mRX),该算法同时考虑了原始线性空间和高维特征空间的异常检测结果,使异常检测效果更加稳定。在仿真图像和真实高光谱图像的实验中,上述2种算法与原始的RX、正则化RX(rRX)和核RX(kRX)3种算法进行了比较,使用了双窗口技术和核主成分分析(KPCA)进行特征提取和基于高阶统计量的特征选择作为预处理来降低数据维数,并在未降维数据上比较上述5种算法。最后,使用ROC曲线评价检测效果,结果表明:提出的2种算法提高了检测效果并具有一定鲁棒性。

Kernel matrix was regularized for improving the stability of the kernel RX algorithm in anomaly detection for hyperspectral images,and regularized kernel RX（rkRX） algorithm was proposed.The RX algorithm was improved by merging the normalized regularized RX algorithm and the regularized kernel RX algorithm,named merging RX algorithm（mRX）.Considering the results of both original linear space and high dimensional feature space at the same time,it produced an improved and more stable performance in anomaly detection.Compared with original RX,regularized RX（rRX） and kernel RX（kRX）,the above two algorithms used double window technique,kernel principal component analysis（KPCA） feature extraction and feature selection based on high order statistics as a preprocessing for reducing data dimension,in simulation images and the real hyperspectral images experiments.Also,the five algorithms were compared in the images data without dimension reduction.Finally,the ROC curves were painted for evaluating the detection performances.The results show that the proposed two algorithms improve the detection performance and have certain robustness.