基于CUDA的影像配准与拼接方法研究

发布时间：2018-07-30 08:33

【摘要】：图像作为客观世界能量或状态的可视化形式,为人们认知和改造客观世界提供了丰富的信息。而图像配准和图像拼接作为图像处理的基本问题之一,在虚拟现实、地质勘测、医学影像、气象预测、应急响应等领域都有着广泛的应用。图形处理器GPU将更多的晶体管用作执行单元,计算能力远远超过传统的中央处理器CPU。目前,GPU技术已经广泛用于数据挖掘、数理统计、图像语音识别、基因工程、全球气候准确预报等领域,同时也为遥感影像的快速处理提供了一种新的解决方案。针对当前影像配准和拼接计算中存在问题,本文在分析当前影像配准和拼接技术的基础上,结合CUDA并行计算技术,重点研究了大尺寸高分辨率遥感影像配准方法以及航空影像的在线实时拼接方法。本文的主要研究内容包括：(1)大尺寸遥感影像配准。针对传统遥感影像配准方法难以适应于大尺寸高分辨率遥感影像配准的问题,本文提出一种由粗到精的配准控制点匹配方法；在此基础上,采用小面元微分纠正方法实现大尺寸影像的高精度纠正,并采用一种自适应的扫描线填充算法来计算每个纠正像素所在的三角形；本文通过分析配准中各个步骤的计算瓶颈问题,利用CUDA并行计算技术对控制点匹配和影像纠正两个阶段进行了加速。通过IKONOS全色影像、Geoeye全色影像和多光谱影像、ZY-3卫星影像等实验表明,本文方法可以取得较高的配准精度,且利用GPU加速算法获得了较高的加速比。(2)航空影像实时在线拼接。针对航空影像在线实时拼接的计算瓶颈,本文提出了一种基于CPU与GPU协同处理的在线拼接方法。基于POS数据,在CPU端计算原始影像与纠正影像之间的单应变换关系,然后利用GPU并行计算实现影像的纠正过程；由于航空影像间具有较大重叠度,因此本文提出一种自适应的拼接方法,即通过计算后续影像的重叠度以判断当前影像是否需要拼接,大大了减少冗余计算；同时本文利用两台计算机进行了模拟实验,其中一台作为模拟相机拍摄影像,往处理机器上传输影像。实验结果表明,该方法基本实现了航空影像的在线实时拼接,且拼接结果满足为灾害、突发情况的实时救援提供决策信息需求。本文研究所提出的配准和拼接方法采用Visual C++进行了实现,并开发了相应原型系统。有关实验结果表明本文方法能实现大尺寸遥感影像的高效配准以及航空影像的在线拼接。文中图26个,表3个,参考文献48篇。
[Abstract]:As a visual form of the energy or state of the objective world, image provides abundant information for people to recognize and transform the objective world. As one of the basic problems of image processing, image registration and image mosaic are widely used in the fields of virtual reality, geological survey, medical image, meteorological prediction, emergency response and so on. The graphics processor GPU uses more transistors as executive units, much more computing power than traditional central processor CPUs. At present, GPU technology has been widely used in the fields of data mining, mathematical statistics, image speech recognition, genetic engineering, accurate prediction of global climate and so on. At the same time, it provides a new solution for the rapid processing of remote sensing images. Aiming at the existing problems in image registration and stitching calculation, this paper analyzes the current image registration and stitching technology, and combines CUDA parallel computing technology. The registration method of large scale and high resolution remote sensing image and the online real-time mosaic method of aerial image are mainly studied in this paper. The main contents of this paper are as follows: (1) large scale remote sensing image registration. Aiming at the problem that the traditional remote sensing image registration method is difficult to adapt to the large scale and high resolution remote sensing image registration, this paper proposes a registration control point matching method from coarse to fine. The small panel differential correction method is used to realize the high accuracy correction of large scale image, and an adaptive scan line filling algorithm is used to calculate the triangle in which each corrected pixel is located. In this paper, by analyzing the bottleneck problem in each step of registration, the CUDA parallel computing technique is used to accelerate the control point matching and image correction. The experiments on IKONOS panchromatic and multispectral images show that the proposed method can achieve high registration accuracy and obtain a high speedup ratio by using GPU acceleration algorithm. (2) Real-time online stitching of aerial images. Aiming at the bottleneck of online real-time mosaic of aerial images, this paper presents an online stitching method based on CPU and GPU. Based on the POS data, the monoclinic transformation between the original image and the corrected image is calculated at the CPU end, and then the correction process of the image is realized by using the GPU parallel computation. In this paper, an adaptive stitching method is proposed, that is, by calculating the overlap degree of subsequent images to determine whether the current images need to be spliced, the redundant computation is greatly reduced, and at the same time, two computers are used to carry out simulation experiments. One of them takes the image as an analog camera and transmits the image to the processing machine. The experimental results show that the method can basically realize the online real-time mosaic of aerial images, and the results of the stitching can provide the decision information for the real-time rescue of disasters and emergencies. In this paper, the proposed registration and splicing methods are implemented with Visual C, and the corresponding prototype system is developed. The experimental results show that the proposed method can achieve the efficient registration of large scale remote sensing images and the online stitching of aerial images. There are 26 figures, 3 tables and 48 references.
【学位授予单位】：中南大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751

【参考文献】