基于稀疏性约束的高光谱图像处理方法研究

发布时间：2018-11-21 08:09

【摘要】：随着传感器技术的发展,光谱成像技术得到了空前的发展。高光谱遥感除了获取图像的空间信息外,还可以得到精细的光谱信息,在军事侦察和国民经济等各个领域的应用越来越广泛。但是,随着高光谱图像的分辨率不断提高,成像光谱仪获取的图像数据已经远远超出了数据传输和处理能力。基于信号稀疏性约束的处理方法近年来广泛应用于信号处理、模式识别和计算机视觉等方面。如何有效的利用光谱图像的稀疏性,已经成为遥感信息处理领域重要的研究方向之一。针对高光谱图像数据处理难题,本文主要分析了高光谱图像的稀疏性,在此基础上研究了基于稀疏性约束的高光谱图像分类和目标检测,主要工作如下:首先,论文分析与验证了高光谱图像的稀疏性。分析高光谱图像数据的典型特性,并利用无监督的学习方法构建字典对高光谱图像数据进行稀疏分解,将图像中实际包含的物理材料的光谱曲线与字典原子进行比对,证明学习的字典原子可以很好地与材料的光谱曲线拟合,验证了高光谱图像的稀疏性。其次,论文提出了基于稀疏嵌入的高光谱图像分类方法。针对高光谱图像的高维特性,利用稀疏嵌入的方法对高光谱图像进行特征提取,通过保持类内紧凑性的条件下进行类内稀疏重建,同时最大限度地增大类间距离,以增强高光谱数据在特征空间投影的离散度。通过对真实数据进行测试表明,本文方法在分类时间和分类精度上比起其它方法都有一定的提高。最后,论文研究了高光谱图像异常检测问题,提出了一种基于金字塔空-谱协同编码的高光谱图像异常检测方法。首先在优化样本-特征分布的行稀疏性、列稀疏性和行分布统计相似性的基础上,采用无监督的学习方法提取低维区分性特征;其次,利用空间金字塔思想在多个空间尺度上对局部像素进行空-谱协同编码;最后统计编码差异性,定位异常。在实测数据集的实验结果验证了方法的有效性和鲁棒性。
[Abstract]:With the development of sensor technology, spectral imaging technology has been unprecedented development. Hyperspectral remote sensing not only can obtain spatial information of images, but also can obtain fine spectral information. It is more and more widely used in military reconnaissance and national economy and other fields. However, with the improvement of the resolution of hyperspectral images, the image data obtained by the imaging spectrometer is far beyond the ability of data transmission and processing. In recent years, signal sparsity constraint based processing methods have been widely used in signal processing, pattern recognition and computer vision. How to make effective use of spectral image sparsity has become one of the important research directions in remote sensing information processing field. Aiming at the difficult problem of hyperspectral image data processing, this paper mainly analyzes the sparsity of hyperspectral image, and then studies the classification and target detection of hyperspectral image based on sparse constraint. The main work is as follows: first, The sparsity of hyperspectral images is analyzed and verified. This paper analyzes the typical characteristics of hyperspectral image data, constructs a dictionary to sparse decompose the hyperspectral image data by using unsupervised learning method, and compares the spectral curve of the physical material actually contained in the image with the dictionary atom. It is proved that the dictionary atoms can fit well with the spectral curves of the materials, and the sparsity of hyperspectral images is verified. Secondly, a method of hyperspectral image classification based on sparse embedding is proposed. In view of the high dimensional characteristics of hyperspectral images, the method of sparse embedding is used to extract the features of hyperspectral images, and the intra-class sparse reconstruction is carried out under the condition of keeping intra-class compactness, and the distance between classes is maximized. In order to enhance the dispersion of hyperspectral data projection in feature space. By testing the real data, it is shown that the classification time and accuracy of this method are better than those of other methods. Finally, the problem of hyperspectral image anomaly detection is studied, and a hyperspectral image anomaly detection method based on pyramid space-spectrum cooperative coding is proposed. Firstly, on the basis of optimizing the row sparsity, column sparsity and the statistical similarity of row distribution, the unsupervised learning method is used to extract the low-dimensional distinguishing features. Secondly, spatial pyramid is used to cocode the local pixels on multiple spatial scales. Finally, the differences of coding are statistically analyzed, and the anomalies are located. The effectiveness and robustness of the method are verified by the experimental results of the measured data sets.
【学位授予单位】：国防科学技术大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP751

【参考文献】