CASIA OpenIR  > 毕业生  > 硕士学位论文
Alternative TitleResearch of Image Matching and Its Application in Retrieval Tasks
Thesis Advisor王春恒
Degree Grantor中国科学院大学
Place of Conferral中国科学院自动化研究所
Degree Discipline模式识别与智能系统
Keyword图像匹配 仿射不变 全局特征 图像检索 计算复杂度 Image Matching Affine-invariant Global Descriptor Image Retrieval Computation Complexity
Abstract图像匹配作为图像处理领域中的基础技术之一,是三维建模、图像检索、目标跟踪等众多高层图像应用的基础,长久以来受到广泛的关注。经过数十年的研究,图像匹配技术得到了长足的进步。目前,经典的图像匹配技术依赖于性能良好的局部描述子和匹配算法,能够较好地处理旋转、尺度和光照等因素的影响。然而,目前的图像匹配算法并不能完全克服仿射变化干扰。 近年来,多种仿射不变的图像匹配算法(ASIFT, Fair-Surf)被提出。上述算法均使用视角模拟的方式保持了局部描述子在不同视角下的一致性,取得了突出的匹配效果。本文在沿用此框架的基础上,引入了全局特征过滤机制,优化了ASIFT算法框架的图像匹配环节,将算法复杂度从 降低为 。同时,本文还对图像匹配技术在图像检索技术中的应用作了一定的探讨,并提出了一种适用于小规模图像库中的精确图像检索算法。本文的主要内容有: 1. 将全局特征过滤机制引入ASIFT匹配算法。ASIFT在匹配环节穷举所有的模拟图像对进行匹配,使得算法复杂度为 ,而其中大部分为非必要计算。本文利用全局特征对图像外观的描述能力,通过计算全局特征相似性过滤掉外观不相似的模拟图像对,降低算法的复杂度至 。 2. 针对本地数据库的精确检索任务,本文提出了一种融合了图像检索技术与图像匹配技术的算法。该算法首先对图像提取全局特征,在本地数据库中检索外观相似的相似子集合。然后在该子集合继续执行图像匹配操作,确定精确的检索结果。实验结果表明,本文提出的算法能够在小规模数据库中较好的实现精确检索。在匹配环节中,本文提出了一种快速的二级匹配算法,显著提高了精确匹配环节的算法效率。 3. 基于上述的精确检索算法,本文完成了原型系统的设计和实现。该系统在实验室自采集的图像数据库中测试了本文提出的精确检索算法性能。
Other AbstractAs one of the fundamental techniques of image processing, image matching is widely used in 3D reconstruction, image retrieval, object tracking and so on. It has attracted lots of attention from computer vision researchers. In the past few decades, much progress has been made in this area. The state-of-art image matching algorithm which is based on excellent local descriptor and epipolar geometry has shown great performance under much circumstance (scale variation, rotation). However, it is not fully invariant to affine transformations. Several affine-invariant image matching algorithms (ASIFT, Fair-Surf) have been proposed in recent years. View simulation is used to overcome affine transformations in these algorithms and achieves great performance. In this paper, a filter mechanism based on global descriptor is introduced into the affine-invariant algorithm and filters most unnecessary matching processes in the original framework. Our approach achieve significant improvement in computation complexity, which is compared to ASIFT’s . In this paper, an algorithm which aims to support accurate image retrieval tasks is proposed. It combines the advantages of image matching techniques and content based image retrieval techniques. The main distributions of this paper are listed as follows: Firstly, a filter mechanism based on global descriptor is introduced into the affine-invariant image matching algorithms. ASIFT exhausts all possible simulated image pairs which are mostly meaningless. Our approach filters most of simulated image pairs which are not similar in appearance by means of computing the distance of global descriptors. The computation complexity of the proposed algortihm is compared to ASIFT’s . Secondly, an algorithm which combines the advantages of image matching techniques and content based image retrieval techniques is proposed in this paper. It can support accurate retrieval tasks in a small image dataset. In the first step, an image set is filtered out by content based image retrieval techniques. Then, the classic image matching algorithm is executed between the query image and every single image in the image set. An accelerate image matching algorithm is also proposed in this paper. Thirdly, based on the algorithm proposed above, a prototype system is implemented. The algorithm was tested in a self-collected image dataset.
Other Identifier201128014628032
Document Type学位论文
Recommended Citation
GB/T 7714
陈文龙. 图像匹配方法及其在检索中的应用[D]. 中国科学院自动化研究所. 中国科学院大学,2014.
Files in This Item:
File Name/Size DocType Version Access License
CASIA_20112801462803(1924KB) 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
[陈文龙]'s Articles
Baidu academic
Similar articles in Baidu academic
[陈文龙]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[陈文龙]'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.