CASIA OpenIR  > 毕业生  > 博士学位论文
Alternative TitleResearch on Fast Image Matching
Thesis Advisor王云宽
Degree Grantor中国科学院研究生院
Place of Conferral中国科学院自动化研究所
Degree Discipline控制理论与控制工程
Keyword实时性 快速模板匹配 简单纹理 复杂纹理 图像扭曲 视角变化 海明距离 快速图像旋转 Real-time Fast Template Matching Untextured Objects Textured Objects Image Deformation Perspective Change Hamming Distance Fast Image Rotation
Abstract随着机器视觉技术发展日益成熟,其已广泛应用于医疗诊断、军事、各类自动检测与控制、遥感、科研、智能机器人、生活等方面,取得了巨大的经济与社会效益。作为机器视觉中的关键技术之一,图像匹配的实时性已成为机器视觉应用可行性的主要瓶颈之一。多年以来,学者们一直以算法的精度及鲁棒性作为研究重点,而对于图像匹配算法的实时性研究较少。因此本文针对匹配算法的实时性问题进行了研究,在保障匹配算法鲁棒性及精度的前提下,较大程度提高匹配算法的整体实时性,论文所完成的主要工作有: (1)模板匹配算法往往需快速获取不同旋转角度下的模板图像,针对该问题,提出了一种基于分块技术的图像快速-高精度旋转算法,该算法与基本图像旋转算法具有相同精度的前提下,显著提高了图像的旋转速度,同时创建Dart模型事先存储旋转仿射变换信息,消除图像旋转算法中的浮点运算。 (2)针对简单纹理图像快速匹配问题,提出了一种基于主特征点的模板匹配算法,该匹配算法对图像扭曲、复杂背景、遮挡及各种干扰等均具有较强的鲁棒性,同时提出基于主特征点的搜索策略、模板映射、基于主方向的图像索引等优化方法,使该算法具有良好的实时性。 (3)针对复杂纹理图像的快速匹配问题,提出了一种基于描述符RIBRIEF的图像快速匹配算法。该算法对图像光照变化、噪声及图像模糊均具有较强的鲁棒性,同时由于采用多种优化方法,如描述符索引与描述符聚类相结合,该算法具有非常理想的实时性。实验表明,通过控制特征点数量,两幅图像(尺寸为640x480,一定视角变化)的匹配时间仅为10ms。 (4)针对复杂纹理图像扭曲问题,提出一种图像区域描述符CHRI,该描述符对图像的旋转具有不变性,同时对图像扭曲具有较强鲁棒性。由于采用了多种优化方法,基于描述符CHRI的匹配算法具有良好的实时性。实验表明,相对于SURF匹配算法,该算法在实时性及图像扭曲鲁棒性方面,均具有更好的性能。
Other AbstractAlong with the machine vision technology increasingly mature, it has been widely used in medical diagnosis, military, automatic detection and control, remote sensing, scientific research, intelligent robot and life etc. As one of key technologies, real-time performance of image matching algorithms has become one of the main considerations for practical application. In recent years, the precision and robustness of matching algorithms have always been the research focal point while the real-time performance hasn’t got enough attention. In this paper, we focus on studying and exploring how to improve real-time performance of matching algorithms. The main work can be summarized as following: (1)To acquire object images of different rotation angles, we proposed a fast-high-quality image rotation algorithm using block division technology. Comparing with basic image rotation algorithm, the runtime of our algorithm decreases significantly under same accuracy, and we proposed Dart model to eliminate float point operation. (2)To realize fast matching of untextured objects, we proposed a fast template matching algorithm based on dominant feature points, which has strong robustness to image deformation, complex background, occlusion and all kinds of disturbances. We proposed some optimization methods such as search strategy based on dominant feature points, image mapping and indexing based on dominant orientation to make algorithm achieve good real-time performance. (3)To realize fast matching of textured objects, we proposed a rotation invariant RIBRIEF descriptor based on BRIEF descriptor, which has strong robustness to illumination, noise and image blur. We proposed some optimization ways such as combining descriptor indexing with descriptor cluster to make algorithm have ideal real-time performance. Experiment results show that the matching speed of two images of size 640x480 is only 10ms by reducing the number of feature points. (4)To realize fast matching of deformed objects, we proposed a kind of image region descriptor CHRI, which is rotation invariant and strongly robust to image deformation. Some optimization methods are used to obtain good real-time performance. Experiment results show that the performance of CHIR is better than SURF in real-time and image distortion.
Other Identifier200818014628011
Document Type学位论文
Recommended Citation
GB/T 7714
李兵. 快速图像匹配算法研究[D]. 中国科学院自动化研究所. 中国科学院研究生院,2011.
Files in This Item:
File Name/Size DocType Version Access License
CASIA_20081801462801(5948KB) 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[李兵]'s Articles
Baidu academic
Similar articles in Baidu academic
[李兵]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[李兵]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.