多平台高光谱图像特征提取适应性的研究

发布时间：2018-02-24 08:00

本文关键词： 高光谱特征提取 CUDA 多平台 GPU 并行计算　出处：《重庆邮电大学》2016年硕士论文　论文类型：学位论文

【摘要】：随着高光谱图像开始呈现出高时间分辨率、高空间分辨率、高光谱分辨率的“三高”特性,数据量也随之出现海量增长的趋势,使其处理起来越为复杂、耗时,对其进行处理的硬件平台环境的要求也大大提高。高光谱图像的近实时、实时处理成为一个难题。目前,国内外的很多学者提出了从一种平台纵向的研究探讨高光谱图像的处理,但几乎都是从单一平台或单方面入手探讨对高光谱图像的处理。而对高光谱图像处理的语言、平台繁多,缺乏将多种语言、平台资源整合、分析比较,以及归类整理。本文正是基于这条思路,从横向和纵向方面针对高光谱图像数据的特点,结合对其处理的语言、平台环境入手,以处理时间、加速比、图像处理效果、能耗、复杂度等方面作为评价指标,探寻一种真正适宜于高光谱图像处理的语言、平台环境。从横向和纵向方面,对高光谱图像处理平台适宜性进行研究提供了一种新思路。本文的主要研究内容如下:首先,对CUDA并行架构的编程模型、存储器模型、处理机制进行了分析,并探讨了基于CUDA程序性能优化问题。其次,从典型的高光谱图像特征提取算法入手,探讨其缺陷和不足,然后在基于CUDA架构的GPU并行环境下对最小噪声分离和主成分分析两种算法,从数据通信、数据分块、存储器访问、协方差计算等方面进行优化和改进,达到对高光谱图像特征提取加速处理的目的。通过仿真实验发现,优化改进后两种算法最高的加速比达到了122倍,优化了处理时间,提升了加速比。最后,针对目前处理高光谱图像的语言、平台环境繁多各异的情形,提出一种多平台机制(ENVI、Matlab、串、并行环境平台)对高光谱图像进行特征提取仿真实验探究,对实验结果从横向方面研究比较,评价各种处理平台的综合优劣,为高光谱图像处理平台的适应性提供一种研究新思路。进而为高光谱图像的分类、目标探测、混合像元分解等后续研究工作的开展打下了基础,为高光谱图像的快速和高效处理带来了可能。
[Abstract]:As hyperspectral images begin to exhibit the "three high" characteristics of high time resolution, high spatial resolution, and high spectral resolution, the amount of data is increasing rapidly, which makes the processing more complex and time-consuming. The requirement of hardware platform for processing is also greatly improved. The near real time and real time processing of hyperspectral image has become a difficult problem. Many scholars at home and abroad have proposed to discuss the processing of hyperspectral images from the longitudinal study of a platform, but almost all of them discuss the processing of hyperspectral images from a single platform or unilaterally, and the language of hyperspectral image processing. There are many platforms, lack of integration of multiple languages, platform resources integration, analysis and comparison, and classification. Based on this idea, this paper aims at the characteristics of hyperspectral image data from the horizontal and vertical aspects, combined with the processing language of hyperspectral image data. The platform environment starts with processing time, speedup ratio, image processing effect, energy consumption, complexity and so on as the evaluation index, explores a kind of language which is really suitable for hyperspectral image processing, platform environment. The main contents of this paper are as follows: firstly, the programming model, memory model and processing mechanism of CUDA parallel architecture are analyzed. The performance optimization problem based on CUDA program is discussed. Secondly, starting with the typical feature extraction algorithm of hyperspectral image, the defects and shortcomings of the algorithm are discussed. Then in the GPU parallel environment based on CUDA architecture, the two algorithms of minimum noise separation and principal component analysis are optimized and improved from the aspects of data communication, data block, memory access, covariance calculation and so on. The simulation results show that the maximum speedup ratio of the two improved algorithms is 122 times, the processing time is optimized, and the speedup ratio is increased. In view of the situation that hyperspectral images are processed in different languages and different platforms at present, a multi-platform mechanism named ENVI Matlab, string and parallel environment is proposed to simulate the feature extraction of hyperspectral images. The experimental results are compared horizontally, and the comprehensive advantages and disadvantages of various processing platforms are evaluated, which provides a new way of research for the adaptability of hyperspectral image processing platform, and then for the classification of hyperspectral images and target detection. The following research work, such as mixed pixel decomposition, has laid a foundation for the rapid and efficient processing of hyperspectral images.
【学位授予单位】：重庆邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP751

【参考文献】