基于图形处理器的SIFT算法研究-空间信息科学与技术专业论文.docxVIP

Abstract Image Registration is an important technique in Image Processing, it provides foundamental support for many advanced Image Processing techniques. Based on different measures, Image Registration methods can be categorized into two major types: Area (Intensity) - based methods and Feature – based methods. Due to the stability in performance, flexibility in implementation and feasibility in application, Feature – based methods have been broadly investigated and implemented. The SIFT algorithm is a popular and stable Feature – based Image Registration method because it has the Scale-Invarient and Rotation-Invarient property, it also has the robustness of toleration to illumination and viewpoint variation, these features made it increasingly hot in academic research and engineering application. Although SIFT algorithm has a good performance, it lacks of quickness and efficiency which has largely limited its usage in an extensive area, such as video search, target tracking, object recognition and other time critical application. To tackle this problem, in order to improve the performance of SIFT in these applications, this paper introduces the CPU+GPU heterogenous computing platform, proposes a heterogenous implementation of SIFT algorithm. It transplants the most time-consuming part of SIFT algorithm Scale Space Generation, Gaussian Convolution and other arithmetically intensive computation to GPU. By taking the advantage of the super scale Stream Processor Array and fast on-die memory, it dramatically improved the speed and efficiency which lays a foundation for further research and application. The general procedure of Feature – based Image Registration can be divided to four stages: Feature extraction, feature space generation, feature matching and image alignment. This paper focuses on feature extraction. Based on a complex platform which comprised of Microsoft Visual Studio 2008 Express, OpenCV 2.1, CUDA 3.2.16, this research conducts investigation and