DSpace About DSpace Software
 

Institutional Repository at Tsinghua University >
本科生优秀毕业论文 >
本科生优秀毕业论文(2012) >

Please use this identifier to cite or link to this item: http://hdl.handle.net/123456789/2491

Title: 基于特征匹配的网络图片聚类与分析
Other Titles: an image clustering algorithm based on feature matching
Authors: 喻川
自动化系
Keywords: 图片聚类
SIFT特征匹配
图片召回率
image clustering
SIFT feature matching
image recall rate.
Issue Date: 20120613
Abstract: 随着社交网站和网络相册应用的兴起,网络图片的自动管理和利用网络图片进行场景的3D重建成为了两个越来越受关注的话题。这两个问题的核心都是如何对图像进行聚类。本文在一个现有的图像聚类算法的基础上,改进了图片匹配算法,完善了聚类算法框架,提高了图片的匹配准确度和成功率,大大地提高了图片召回率。匹配算法的改进思路如下,首先通过两幅图片的初步匹配得到其对极几何模型,然后用该模型反馈验证这两幅图片所有的特征点对,从而得到更多准确的特征点对。同时本文还利用匹配结果估计图片主体区域的SIFT特征点集,从而在保证丢失有效特征点不多的前提下大大减少特征点数目,加快匹配速度。最后本文引入模糊聚类的思想来改进算法框架,即将数据集中每张图片和多个子类的代表图片进行匹配,而不只是将其和“最像”的一幅代表图片进行匹配,从而找回更多的图片。 本文利用上述聚类算法实现了一个网络图片的分类和识别系统,该系统有以下几个功能:将不同场景的图片按其所含内容分成若干类,将同一场景的图片按其特性(如拍摄角度,拍摄远近等)分成多个子类,并剔除其中的干扰图片,判断输入的图片是否含有已知的场景等。
With the growth of social networking sites and web album applications, the automatic management of network pictures and 3D reconstruction of landmarks, based on pictures gathered from the Internet, has become two increasingly hot topics. The cores of these two issues are how to cluster the images. Based on an existing and mature image clustering algorithm, this paper presents a better approach of image clustering that has a higher picture recall rate. Our main work is proposing a better SIFT feature matching algorithm and improving the algorithm framework.To improve the SIFT feature matching algorithm, we use the fundamental matrix of the two pictures to perform verification of all the potential SIFT pairs of the two pictures and to get more SIFT pairs.Then we use the matching results between one picture and some of other pictures to estimate its main aero.Then by introduce the idea of fuzzy clustering, we match each picture with several representatives instead of the only “nearest” one. After this improvement of the algorithm framework, we can get a much higher recall rate.
Description: 14
URI: http://hdl.handle.net/123456789/2491
Appears in Collections:本科生优秀毕业论文(2012)

Files in This Item:

File SizeFormat
025喻川2008011544.pdf3.95 MBAdobe PDFView/Open

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

 

Valid XHTML 1.0! DSpace Software Copyright © 2002-2010  Duraspace - Feedback