提出一种基于低秩表示和学习字典的高光谱遥感图像异常探测算法。相对于其它低秩矩阵分解方法如鲁棒主成分分析, 低秩表示方法更为契合高光谱图像的线性混合模型。该算法将低秩表示模型应用到高光谱图像异常探测问题上来, 引入表征背景信息的学习字典, 大大增强了低秩表示模型对初始参数的鲁棒性。仿真和实际高光谱数据的实验结果表明, 所提出的算法有效地提高了异常的探测率, 同时对初始参数具有较好的鲁棒性, 可以作为一种解决高光谱图像异常探测的有效手段。

This paper proposes an anomaly detection method based on low-rank representation and learned dictionary for hyperspectral imagery. The model of low-rank representation, which fits the linear mixing model of hyperspectral imagery more precisely compared with other low-rank decomposition algorithms such as robust principle component analysis (RPCA), was introduced to settle the anomaly detection problem for hyperspectral imagery. To improve its robustness to initialized parameters, a learned dictionary that represents only background information was adopted in the proposed method. Experiments on synthetic and real hyperspectral datasets illustrated that the proposed method is capable of improving detection results. Meanwhile, it is robust to initialized parameters and can be viewed as an effective technique to detect anomalies in hyperspectral imagery.