基于最小噪声分离变换的高光谱异常检测方法研究

RX算法和核RX算法能很好地分离目标和背景,是较为广泛使用的异常检测算法,但是高光谱图像数据量大且存在冗余信息和噪声,直接进行RX及核RX异常探测运算量大且容易受噪声影响.针对此问题,提出一种基于最小噪声分离变换的高光谱图像异常检测方法,首先采用残差分析法估计噪声协方差矩阵以改进最小噪声分离变换,然后利用改进后的最小噪声分离变换来有效地降低高光谱图像数据的维数并分离出噪声,最后对低维降噪后的数据进行RX及核RX异常检测,避免了随机噪声对RX及核RX异常检测结果的影响并提高了异常检测率.对真实的AVIRIS数据测试表明,该算法优于传统的相应的RX、核RX异常检测算法.

ON ANOMALY DETECTION OF HYPERSPECTRAL IMAGE BASED ON MINIMUM NOISE FRACTION

The RX and kernel RX algorithms are widely used in anomaly detection for their improvement in the separation between target and background pixels.However,the huge data and redundant noise of hyperspectral image data make it difficult to apply the RX or kernel RX anomaly detection directly due to heavy computation load and susceptibility on noise impact.To solve this problem,this paper proposes a novel anomaly detection algorithm for hyperspectral images based on minimum noise fraction(MNF).Firstly,we use residual analysis method for noise covariance matrix estimation to improve the MNF.Secondly,the improved MNF is used to reduce the dimension of hyperspectral image data and to separate the noise from signals effectively.Finally,the RX and kernel RX anomaly detections are implemented on low-dimensional denoised data.In this way,the detrimental effect of random noise on the RX and kernel RX anomaly detection results is avoided,and the anomaly detection rate is increased.Test on actual AVIRIS data shows that the algorithm proposed in the paper outperforms the corresponding traditional RX and kernel RX anomaly detection algorithm.