基于多特征信息的低秩表示高光谱影像异常探测

发布时间：2018-04-09 22:20

本文选题：高光谱遥感影像　切入点：异常探测　出处：《中国矿业大学》2017年硕士论文

【摘要】：高光谱遥感影像具有丰富的空间与光谱信息,利用高光谱遥感影像进行异常探测成为目标探测领域的研究重点。为提升高光谱影像异常探测精度,论文对基于低秩表示的高光谱异常探测模型和特征提取进行研究和改进。现有基于低秩表示的高光谱异常探测模型存在不足,一方面仅利用光谱信息,并没有考虑像素空间邻域信息;另一方面字典对探测结果的稳定性造成影响。因此,论文主要在空-谱信息结合、字典构造方向进行研究;同时对隐式低秩表示特征提取算法进行改进,并将其应用到高光谱影像异常探测中。论文通过研究高光谱遥感影像的空间约束构建、字典学习及特征提取等三个方面,并利用合成数据与真实实验数据验证算法在高光谱影像异常探测中的有效性。论文的主要研究内容如下:(1)在基于低秩表示的高光谱遥感影像异常探测模型中构建单局部窗口,建立中心测试像素低秩表示系数与邻域像素表示系数的空间约束关系。鉴于单局部窗口的局限性,将其扩展为多局部窗口,减少局部窗口邻域中出现异常像素的概率,提升异常探测精度。(2)基于高光谱影像数据的全局低秩性特点,在稀疏字典模型基础上,添加字典低秩约束项,建立低秩字典学习模型,增强字典对低秩表示高光谱影像异常探测的稳定性。(3)在高光谱影像特征提取研究中引入隐式低秩表示方法,通过对表示系数添加稀疏约束项对其改进,构建新的特征空间,并在该特征空间中采用经典探测算法,提升异常探测精度。
[Abstract]:Hyperspectral remote sensing images are rich in spatial and spectral information, and the use of hyperspectral remote sensing images to detect anomalies has become the focus of research in the field of target detection.In order to improve the accuracy of hyperspectral anomaly detection, the model and feature extraction of hyperspectral anomaly detection based on low rank representation are studied and improved in this paper.The existing hyperspectral anomaly detection model based on low rank representation is insufficient. On the one hand, only spectral information is used, and pixel spatial neighborhood information is not considered; on the other hand, dictionary has an effect on the stability of detection results.Therefore, this paper mainly combines space-spectrum information, dictionary construction direction, and improves the implicit low-rank representation feature extraction algorithm, and applies it to hyperspectral image anomaly detection.In this paper, the spatial constraint construction, dictionary learning and feature extraction of hyperspectral remote sensing images are studied, and the effectiveness of the algorithm in detecting hyperspectral image anomalies is verified by using synthetic data and real experimental data.The main contents of this paper are as follows: (1) A single local window is constructed in the hyperspectral remote sensing image anomaly detection model based on low rank representation, and the spatial constraint relationship between the central test pixel low rank representation coefficient and the neighborhood pixel representation coefficient is established.In view of the limitation of single local window, we extend it to multi-local window, reduce the probability of abnormal pixels appearing in the neighborhood of local window, and improve the detection accuracy of anomaly. (2) based on the global low rank characteristic of hyperspectral image data,On the basis of sparse dictionary model, the low rank dictionary learning model is established by adding dictionary low rank constraint items.The stability of enhanced dictionary for anomaly detection in low-rank hyperspectral images is analyzed. (3) an implicit low-rank representation method is introduced into the study of hyperspectral image feature extraction, and a new feature space is constructed by adding sparse constraint items to the representation coefficients.In this feature space, the classical detection algorithm is used to improve the accuracy of anomaly detection.
【学位授予单位】：中国矿业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：P237

【参考文献】