基于GPU的遥感图像配准并行算法研究及应用系统实现

发布时间：2018-03-11 04:24

  本文选题：遥感图像　切入点：配准　出处：《国防科学技术大学》2014年硕士论文　论文类型：学位论文

【摘要】：图像配准在许多遥感应用中是一个重要的、不可缺少的步骤。遥感图像的规模随着数据分辨率的不断提高而日渐增大;同时,图像配准是一个典型的计算和访存密集型过程,计算复杂度较高,采用传统串行处理模式已无法满足军事、农林等高端应用的实时性处理需求。随着GPU计算性能和可编程性的不断提升,GPU通用计算已成为计算机技术领域的研究热点,这为加快遥感图像的处理速度提供了新的思路。本文针对基于区域和基于特征两类配准中的两种典型方法,深入研究了基于GPU的遥感图像配准并行算法及优化策略,并面向实际应用设计实现了相应的并行处理软件原型系统。本文的主要工作和贡献体现在以下几个方面:1.研究理解了CPU-GPU异构执行模式。研究了以n VIDIA公司GPU为代表的GPU体系结构和相应的CUDA编程模型,系统掌握了使用CPU-GPU异构模式开发并行算法的基本技能。2.研究并提出了基于GPU的遥感图像全局配准并行算法。选取一种基于相关系数全局配准算法作为GPU并行算法设计和优化的基础,给出了适合该类方法的GPU并行设计,并从数据加载、线程访存、通信与同步等几个方面给出了针对性的优化实现策略。实验结果表明,GPU并行程序获得了良好的性能加速比。3.研究并提出了基于GPU的遥感图像控制点匹配并行算法。搜索控制点和基于控制点的匹配参数计算是该类配准方法的核心步骤,该步骤涉及不规则数据访问、多重分支、循环迭代等数据相关问题,并行设计和优化更为困难。选取一种基于互信息的控制点匹配算法作为研究对象,在数据流分析的基础上,重点针对互信息计算和最小二乘匹配过程设计了两种GPU并行实现方案。实验结果表明,在难以消除迭代相关的情况下,通过优化利用本地存储、原子操作等方法使得GPU程序仍然获得了10倍以上的加速效果。4.设计实现了一个基于Web的遥感图像并行处理原型系统。系统采用B/S模式,基于Java语言开发,在Spring、Hibernate、Struts框架基础上提供图像处理服务,集成了包括上述配准算法研究成果在内12类共49种遥感图像并行处理算法。系统提供了友好的交互界面并具有良好的可扩展性。
[Abstract]:Image registration is an important and indispensable step in many remote sensing applications. The scale of remote sensing images increases with the increasing resolution of data. At the same time, image registration is a typical computation- and memory-intensive process. Because of the high computational complexity, the traditional serial processing mode can no longer meet the military requirements. With the development of GPU computing performance and programmability, it has become a research hotspot in the field of computer technology. This provides a new way of thinking to speed up the processing of remote sensing image. In this paper, the parallel algorithm and optimization strategy of remote sensing image registration based on GPU are studied, aiming at the two typical methods of register-based and feature-based registration. The corresponding parallel processing software prototype system is designed and implemented for practical applications. The main work and contributions of this paper are as follows: 1.The heterogeneous execution mode of CPU-GPU is studied and understood, and the GPU of n VIDIA company is studied. Table GPU architecture and corresponding CUDA programming model, The system has mastered the basic skill of developing parallel algorithm using CPU-GPU heterogeneous mode. 2. The global registration parallel algorithm of remote sensing image based on GPU is studied and proposed. A global registration algorithm based on correlation coefficient is selected as GPU parallel algorithm. The basis of design and optimization, The GPU concurrent design suitable for this kind of method is given, and the data is loaded, and the memory is accessed by thread. Several aspects such as communication and synchronization are given. The experimental results show that the parallel program achieves a good speedup ratio. 3. The parallel algorithm of remote sensing image control point matching based on GPU is studied and proposed. Searching control points and calculating matching parameters based on control points are the core steps of this kind of registration method, This step involves irregular data access, multiple branches, cyclic iteration and other data-related problems. It is more difficult to design and optimize in parallel. A control point matching algorithm based on mutual information is selected as the research object, based on the data flow analysis. Two parallel implementation schemes of GPU are designed for mutual information computation and least square matching. The experimental results show that the local storage can be optimized under the condition that iterative correlation is difficult to be eliminated. Atomic operation and other methods make the GPU program still get more than 10 times the acceleration effect. 4. A prototype system of remote sensing image parallel processing based on Web is designed and implemented. The system adopts B / S mode and is developed based on Java language. Based on Spring hibernate Struts framework, image processing services are provided, and 49 parallel remote sensing image processing algorithms are integrated into 12 kinds of remote sensing image parallel processing algorithms, including the results of registration algorithms mentioned above. The system provides a friendly interactive interface and has good scalability.
【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】：2014
【分类号】：TP751
，

本文编号：1596566