CASIA OpenIR  > 毕业生  > 博士学位论文
鲁棒的局部图像特征匹配方法
Alternative TitleRobust Local Image Feature Matching
王振华
Subtype工学博士
Thesis Advisor吴福朝
2014-05-17
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特征检测 特征描述 特征匹配 光照不变性 仿射不变性 Feature Detection Feature Description Feature Matching Illumination Invariance Affine Invariance
Abstract图像的特征匹配是计算机视觉研究中最基本的问题之一,也是很多视觉应用的基本步骤,如物体识别、三维重建、全景图拼接、物体跟踪等等。长期以来,图像特征匹配受到了计算机视觉学术界研究人员的广泛关注。图像的特征匹配基本思想都是首先在图像中检测出不变的特征点或者特征区域,然后为其计算鲁棒、高辨识能力的特征描述子,最后在某种相似性度量下建立特征之间的匹配关系。由于成像条件的差异性,几何、光照变化的复杂性,图像内容的多样性等,图像特征匹配方面的研究仍然面临着很多悬而未决的问题和挑战。本文针对这些问题进行了较为深入的研究,主要贡献和创新点包括以下三个方面: (1) 针对实时快速的特征匹配问题,提出了一种快速鲁棒的不变性特征(FRIF,Fast Robust Invariant Feature)。该特征包括基于快速近似的拉普拉斯-高斯(FALoG,Fast Approximated Laplacian of Gaussian)算子的特征点检测算法和基于混合信息的二进制描述子两部分。FALoG能够在尺度空间中计算稳定的尺度不变特征点。由于计算中使用了滤波器分解法和积分图像法,FALoG极大的提高了运算速度。基于混合信息的二进制描述子综合考虑了局部面片内采样点的梯度信息和采样点之间的灰度大小关系,提高了传统二进制描述子的辨识能力。同时,二进制表达形式不仅存储空间小,而且能够利用海明距离快速计算匹配。 (2) 针对仿射形变下的鲁棒特征匹配问题,提出了一种仿射变换参数化下基于子空间表示的特征描述子(ASR,Affine Subspace Representation)。和传统处理仿射形变的方法相比,本方法无需在底层采用仿射协变的特征区域检测算法,而是在描述子的构造过程中考虑仿射不变性。具体而言,首先通过一种称为PCA-patch向量的紧凑中间表示来描述不同视角下局部面片;然后,基于子空间假设,使用低维的线性子空间来表示同一特征点在不同视角下的PCA-patch向量集合;最后,通过采用一种子空间到点的映射,将线性子空间表示转换为向量形式存储,得到ASR描述子。为加快运行时速度,本方法还引入了一种快速近似算法。该算法将面片的映射和面片的中间表示合并为一个过程,并将大量的计算转移到离线学习的过程中,因此能极大的提高描述子的计算速度。由于本方法在构造描述子时对多个视角下的信息进行了建模表示,因此具备了较强的区分能力,并且能处理一定程度的仿射形变。 (3) 针对复杂光照变化下的鲁棒特征匹配问题,提出一种基于全局和局部灰度序信息的特征描述方法。首先,建立两种特征:局部灰度序模式(LIOP,Local Intensity Order Pattern)和全局灰度序模式(GIOP,Global Intensity Order Pattern)。其中,LIOP用于表示特征点局部面片内每个像素点周围的采样点之间的局部灰度序信息,而GIOP用于表示特征点局部面片内每个像素点周围的采样点粗略的全局灰度序信息。然后,通过在每个灰度序区域内分别统计LIOP和GIOP两种局部特征,构造LIOP和GIOP描述子。进一步的观察发现,LIOP和GIOP刻画的灰度序信息具有很强的互补性,从而提出了一种基于混合灰度序模式(MIOP,Mixed Intensity Order Pattern)的特征描述子,它显著的提...
Other AbstractEstablishing feature correspondences is a fundamental step for many computer vision applications, such as object recognition, 3D reconstruction, panoramic image stitching and object tracking. Thus, it has received many attentions in the community of computer vision. The basic idea is to first detect interest points or interest regions and then compute robust and distinctive feature descriptors on each of them. Once the feature descriptors are computed, the feature correspondences between different images can be automatically established under some similarity measure. Due to the different imaging conditions, the complexities of geometric and photometric transformations, and the variety of scene types, there are still lots of challenging problems lying in this area. This dissertation aims to deal with these problems in local image feature matching. The main contributions and novelties can be summarized as follows: (1) To obtain a high quality feature while maintaining a low computational cost, a novel Fast Robust Invariant Feature (FRIF) is proposed for both feature detection and description. For the detection part, a Fast Approximated Laplacian-of-Gaussian (FALoG) detector is proposed to select scale-invariant keypoints. By using factorization and integral image techniques, the responses of FALoG could be computed very fast. For the description part, a mixed binary descriptor is proposed to encode both local pattern and inter-pattern information, which improves the distinctiveness of traditional binary descriptors. As the descriptor is indeed a binary string, it only takes very small spaces to be stored. The correspondences also can be established very fast by using Hamming distance. (2) To deal with affine distortion, a novel Affine Subspace Representation (ASR) descriptor is proposed. Unlike the traditional methods which rely on an underlying affine covariant region detector to deal with affine distortion, ASR removes such a dependency, and considers the affine invariance at the description stage. Specifically, a set of PCA-Patch coefficients vectors are first computed to encode the local information of warped patches for their compactness and efficiency. Base on the subspace assumption, the PCA-Patch coefficients vectors under various affine transformations of the same point are represented by a low dimensional linear subspace, and the ASR descriptor is obtained by using a simple subspace to point mapping. By introducing a fast approximate algorit...
shelfnumXWLW1986
Other Identifier201018014628063
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/6578
Collection毕业生_博士学位论文
Recommended Citation
GB/T 7714
王振华. 鲁棒的局部图像特征匹配方法[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_20101801462806(5026KB) 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[王振华]'s Articles
Baidu academic
Similar articles in Baidu academic
[王振华]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[王振华]'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.