融合PSO的N-FINDR改进端元提取算法研究

发布时间：2019-05-11 23:18

【摘要】：随着成像和处理技术的进步,高光谱遥感在地质勘探、军事应用、植被检测、海洋遥感等领域发挥着越来越重要的作用。但是由于仪器空间分辨率的局限性和地球表面结构的复杂多样性,图像中的一个像元往往包含着多种不同的地物类型,从而形成了混合像元。 “端元”是高光谱图像中能详尽表示待测地物光谱属性的纯像素,可以作为后续高光谱图像处理算法的先验知识。获取能够很好地反映待研究地物光谱属性信息的端元向量,是对高光谱数据做进一步分析的重要前提。N-FINDR是一个经典且有效的端元寻找算法,能够在保证丰度约束性的同时,获取图像的实际像元作为端元,对后续的分类、识别和解混等作用显著。但理想的N_FINDR算法需要遍历所有可能的像元组合,计算量巨大。而目前用于加快搜索速度的算法,其最终结果大多在一定程度上受到样本排序和初始端元集选择的影响。另外,目前所关注的端元主要包括如军事目标探测中的异常点,以及图像组分分析中的大面积成分端元点。在进行图像主要组分分析时,过多无关端元的参与会降低解混和分类的精度,但目前的端元寻找算法大多并未对这两类端元进行区分。论文的研究工作主要包括以下三部分：第一,初始端元集的优化。利用基于相关性分析的N-FINDR算法获取端元集作为初始端元集,降低初始端元间的关系及其对最终结果的不利影响；第二,利用粒子群算法进一步优化候选端元。对于与初始端元相关性系数大于某一阈值的所有像元向量进行粒子群优化,以保证选出更接近于真实端元的像元作为最终端元。第三,利用端元变异性定义粒子群算法的优化目标。定义新的目标函数为以每个候选端元为类别中心的Fishier比,即以单形体体积作为类间变异,阈值内像元的方差为类内变异。选择能够最大化该Fishier比值的像元为端元,实现了对异常点(包括噪声点)的抑制。最后,利用模拟数据生成混合像元的影像,验证改进N-FINDR算法的有效性。以海上溢油检测和分析为问题背景,关心的主要目标组分是油、水和船。利用真实的机载海上溢油图像对论文中的算法进行测试,结果进一步验证了论文所提出算法的有效性。
[Abstract]:With the development of imaging and processing technology, hyperspectral remote sensing plays a more and more important role in geological exploration, military applications, vegetation detection, marine remote sensing and other fields. However, due to the limitation of the spatial resolution of the instrument and the complex diversity of the earth's surface structure, a pixel in the image often contains many different types of ground objects, thus forming a mixed pixel. "end element" is a pure pixel in hyperspectral image, which can represent the spectral properties of the object to be tested in detail, and can be used as a priori knowledge of the subsequent hyperspectral image processing algorithm. Obtaining the end element vector which can well reflect the spectral attribute information of the ground object to be studied is an important prerequisite for further analysis of hyperspectral data. N-FINDR is a classical and effective end element search algorithm. It can not only ensure the abundance constraint, but also obtain the actual pixel of the image as the end element, which plays an important role in subsequent classification, recognition and unmixing. However, the ideal N_FINDR algorithm needs to traverse all possible pixel combinations, and the amount of computation is huge. At present, the algorithm used to speed up the search speed, the final results are mostly affected by the sample sorting and the selection of the initial set of terminal elements to a certain extent. In addition, the end elements concerned at present mainly include abnormal points in military target detection and large area component end points in image component analysis. In the analysis of the main components of the image, the participation of too many unrelated end elements will reduce the accuracy of unmixing and classification, but most of the current end element search algorithms do not distinguish between the two kinds of end elements. The research work of this paper mainly includes the following three parts: first, the optimization of the initial set of elements. The N-FINDR algorithm based on correlation analysis is used to obtain the end element set as the initial end element set to reduce the relationship between the initial end elements and their adverse effects on the final result. Secondly, the particle swarm optimization algorithm is used to further optimize the candidate terminal elements. Particle swarm optimization is carried out for all pixel vectors whose correlation coefficient with the initial end element is greater than a certain threshold in order to ensure that the pixel which is closer to the real end element is selected as the final end element. Third, the end element variability is used to define the optimization objective of particle swarm optimization algorithm. The new objective function is defined as the Fishier ratio with each candidate terminal element as the category center, that is, the volume of the single body is taken as the inter-class variation, and the variance of the pixel within the threshold is the intra-class variation. The pixel which can maximize the Fishier ratio is selected as the end element, and the outliers (including noise points) are suppressed. Finally, the simulation data are used to generate mixed pixel images to verify the effectiveness of the improved N-FINDR algorithm. Based on the detection and analysis of offshore oil spills, the main target components are oil, water and ships. The real airborne offshore oil spill image is used to test the algorithm in this paper, and the results further verify the effectiveness of the proposed algorithm.
【学位授予单位】：大连海事大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751

【参考文献】