SIA OpenIR  > 光电信息技术研究室
Alternative TitleStudy on Image Feature Extraction for Wide Baseline Matching
Thesis Advisor尹健 ; 史泽林
Keyword宽基线 仿射不变 图像特征匹配 距离直方图 多模型
Call NumberTP391.41/L35/2012
Degree Discipline模式识别与智能系统
Degree Name博士
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract本文以宽基线条件下的自动目标识别为研究背景,研究如何提取和匹配仿射不变的局部图像特征,使具有较大基线变化的图片之间仍有较多的匹配特征。 Harris-Affine和Hessian-Affine算法是根据仿射尺度空间理论构造的仿射不变特征提取算法,但事实上其算法性能受到基线宽度的限制,不能适应较宽的基线变化。针对这一不足,提出了以二阶矩矩阵和Hessian矩阵度量特征区域的仿射变形,将特征区域由各向异性转化为各向同性,再使用SIFT算法提取仿射不变特征的改进算法。实验表明,该算法提取的匹配特征数量明显多于Harris-Affine和Hessian-Affine算法。 针对MSER特征提取算法提取的特征数量较少、特征难以描述和匹配的问题,提出了一种完全仿射不变特征提取算法,该算法充分利用MSER特征定位精度和重复率高的优点,根据MSER特征选取变换区域,在变换区域上提取仿射变特征。实验数据表明,该算法与其他现有的算法相比,所提取的仿射不变特征数量更多,验证了算法的有效性。 研究了仿射变形形状归一化及匹配问题。提出了一种确定主方向的新方法,大大简化了形状归一化的数学计算。提出了新的基于距离直方图的形状描述符,实现了归一化形状的匹配。实验验证了算法的有效性。 为了滤除特征匹配中的错误匹配,根据平面单应变换模型验证每个匹配的正确性。PEARL算法是一种有效的多模型验证算法,但它仍有初始模型不准确和类间平滑项设置不合理的问题。针对模型初始化问题,使用SPRT算法对抽样生成的模型进行检验,选取质量较高的模型作为PEARL算法的初始模型,使算法收敛到正确的模型。为PEARL算法的全局能量函数的类间平滑项设定距离阈值,提高模型计算的精度。实验结果表明,改进后的算法计算的模型更加准确和合理。
Other AbstractBased on the research background of automatic target recognition, this dissertation aims at extracting affine invariant local features that can be matched between images which have large baseline variations. Harris-Affine and Hessian-Affine are affine invariant feature extractors which are constructed based on affine scale space. But they are not fit for wide baseline variation as their performances are restricted by the baseline. To solve this problem, an improved algorithm has been proposed. This algorithm uses second moment matrix or Hessian matrix to measure the affine deformation of a feature region. Anisotropic feature regions will be transformed into isotropic according to affine deformations. Finally, affine invariant features will be detected in isotropic image regions by SIFT. Experimental results show that this novel algorithm extracts significantly more features which can be matched in wide baseline cases than Harris-Affine and Hessian-Affine. MSER features are difficult to describe and match. Usually, its number is too little. To solve these problems, a kind of fully affine invariant feature extractor has been proposed. Because MSER features have high repeatability score and can be localized accurately, fully affine invariant feature extractor selectes transform regions based on MSER features. Affine invariant features are extracted in transformed regions. Experimental results have verified the effectiveness of this novel algorithm. It extracts more affine invariant features that can be matched in wide baseline cases compared with other algorithms. Study on the problem of how to normalize and match affine deformed shapes. A novel method has been proposed to determine the dominant direction of the shape which greatly simplifies the shape normalization. A shape descriptor which will be used to match normalized shapes has been constructed by the distance histogram. Experimental results have verified the validity of the algorithm. In order to eliminate incorrect matches, the dissertation use homography to verify every match. PEARL is an effective multi model verification algorithm. But its initial models are not accurate and the inter model smoothing item is unreasonable. Using the SPRT algorithm to verify the initial models which have been calculated from the sampled matches makes PEARL converge to correct models. Set a distance threshold for the inter model smoothing item of PEARL’s global energy function to improve the precision of models. Experimental results show that models calculated by improved algorithm are more accurate and reasonable.
Contribution Rank1
Document Type学位论文
Recommended Citation
GB/T 7714
李威. 面向宽基线匹配的图像特征提取方法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2012.
Files in This Item:
File Name/Size DocType Version Access License
面向宽基线匹配的图像特征提取方法研究.p(6591KB) 开放获取LicenseApplication Full Text
Related Services
Recommend this item
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[李威]'s Articles
Baidu academic
Similar articles in Baidu academic
[李威]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[李威]'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.