JSM-MI算法：解决图像配准中的离群值与局部极大值问题
125 浏览量
更新于2024-08-30
收藏 898KB PDF 举报
本文主要探讨了在图像配准领域中处理离群值图像的问题。传统的互信息（MI）作为一种广泛使用的相似性度量，通过优化联合直方图中簇的紧密度来实现图像的精确配准。然而，MI在面对一个图像中存在的“离群值”对象，即那些在另一图像中不存在的对象时，会表现出显著的敏感性。此外，MI还受到局部最大值和有偏估计的困扰，这可能影响配准的准确性和稳定性。
作者提出了联合显着图（JSM，Joint Saliency Map）作为解决这些问题的新方法。JSM的创新之处在于它能够强调两个图像中的对应显着结构，这些结构被特意组织进加权联合直方图中的平滑且紧凑的簇中。这种方法有效地解决了离群值对象带来的挑战，并通过平滑处理减少了局部最大值的影响，从而提高了配准算法的鲁棒性。
在实验中，JSM-MI（基于JSM的互信息）算法展现出显著的优势。它不仅在常规图像配准任务中表现出高精度，而且在面对含有离群值的困难图像对时，其性能依然保持稳定。因此，这项研究不仅提供了改进的图像配准策略，而且对于那些包含异常情况的图像处理应用具有重要的实际价值。
这篇研究论文关注的是图像配准中的问题，并通过引入JSM来增强对离群值的抵抗能力，同时改善了配准过程中的局部最优问题。这对于提高图像处理系统的稳健性和准确性具有重要的理论和实践意义，特别是在处理复杂、多变的图像数据集时。
IEEE SIGNAL PROCESSING LETTERS, VOL. 17, NO. 1, JANUARY 2010 91
Registration of Images With Outliers
Using Joint Saliency Map
Binjie Qin, Member, IEEE, Zhijun Gu, Xianjun Sun, and Yisong Lv
Abstract—Mutual information (MI) is a popular similarity
measure for image registration, whereby good registration can
be achieved by maximizing the compactness of the clusters in the
joint histogram. However, MI is sensitive to the “outlier” objects
that appear in one image but not the other, and also suffers from
local and biased maxima. We propose a novel joint saliency map
(JSM) to highlight the corresponding salient structures in the
two images, and emphatically group those salient structures into
the smoothed compact clusters in the weighted joint histogram.
This strategy could solve both the outlier and the local maxima
problems. Experimental results show that the JSM-MI based
algorithm is not only accurate but also robust for registration of
challenging image pairs with outliers.
Index Terms—Image registration, joint saliency map, mutual in-
formation, outliers, weighted joint histogram.
I. INTRODUCTION
I
MAGE registration can be considered as finding the op-
timal transformation
between the reference image and
the floating image to maximize a defined similarity measure
such as mutual information (MI). Since 1995 [1], [2], MI has
been proved to be very effective in image registration. The MI
between
and (with intensity bins and ) is defined as:
(1)
where
and
are the entropy of the intensities
of image
and the entropy of the joint intensities of two im-
ages,
is the intensity probabilities with
and , is the joint intensity probabili-
ties estimated by the joint histogram
.
MI-based registration methods take advantage of the fact that
properly registered images usually correspond to compactly-
clustered joint histograms [3]. They measure the joint histogram
dispersion by computing the entropy of the joint intensity prob-
abilities. When the images become misregistered, the compact
Manuscript received August 08, 2009; revised September 27, 2009. First
published October 06, 2009; current version published October 28, 2009. This
work was supported in part by the NSFC (60872102), NBRPC (973 Program
2010CB834303), the Science Foundation of Shanghai Municipal Science &
Technology Commission (04JC14060), Shanghai Municipal Health Bureau
(2008115), and the Small Animal Imaging Project (06-545). The associate ed-
itor coordinating the review of this manuscript and approving it for publication
was Prof. H. Vicky Zhao.
B. Qin, Z. Gu, and X. Sun are with the Department of Biomedical Engi-
neering, School of Life Sciences & Biotechnology, Shanghai Jiao Tong Univer-
sity, Shanghai 200240, China (e-mail: bjqin@sjtu.edu.cn; gzj0126@gmail.com;
sxj_sun@sjtu.edu.cn).
Y. Lv is with the Department of Mathematics, Shanghai Jiao Tong University,
Shanghai 200240, China (e-mail: yslv@sjtu.edu.cn).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/LSP.2009.2033728
Fig. 1. (a)–(b) Intra-operative and pre-operative MR image with a large tumor
resection. (c) Joint histogram dispersion with two clotted clusters (dark red in
pseudo color). (d) Joint saliency map for (a) and (b).
clusters become disperse sets of points in the joint histogram and
the entropy of the joint intensity probabilities increases. Making
no assumptions about the form of the intensity mapping between
the two images, MI is sensitive to the unmatchable outliers, e.g.
the tumor resection in the intra- and pre-operative brain im-
ages [see Fig. 1(a)–(b)]. To reject the outliers, some approaches
are proposed including consistency test [4], intensity transfor-
mation [5], gradient-based asymmetric multifeature MI [6] and
graph-based multifeature MI [7]. However, all these methods do
not emphasize the corresponding salient structures in the two
images to suppress the outliers. Furthermore, MI likely suffers
from local and biased maxima [8] which are caused by the am-
biguities in defining structure correspondence.
Spatial information, i.e. the dependence of the intensities of
neighboring pixels, has been included in MI [9]–[12] to im-
prove registration. Nevertheless, almost all MI-based methods
equally treat each overlapping pixel pair as a separate point in
the overlap area to calculate the joint histogram. This could
raise three issues: 1) when we equally consider the outlier
pixel pairs, the noncorresponding structures overlap and the
histogram will show certain clusters for the grey values of
the outliers. These clusters easily introduce the histogram
dispersion [see Fig. 1(c)] with increasing misregistration; 2)
while registration can be achieved by maximizing the com-
pactness of the histogram, the undesired clotted clusters [see
Fig. 1(c)] related to many noisy pixel pairs in the structureless
regions, such as background and white matter in the brain
image, increase the MI ambiguities and the local maxima [8]
(Fig. 5(c) shows that the normalized MI [1], [20] is in a biased
global maximum when the whole background areas in the two
endoscopic images are exactly aligned); 3) when we group
the intensity pairs as separate points into the histogram, the
independence of the neighboring bins could increase the MI
ambiguities and the local maxima. To solve this problem, joint
histogram smoothing (or blurring) [5], [8] has been used to
increase the dependence of the neighboring histogram bins. We
address these issues above as follows.
In fact, image registration is to match the corresponding
salient structures in both images. To suppress the outliers and
the homogeneous pixel pairs, the corresponding pixel pairs in
the corresponding salient structures should contribute more to
the joint histogram. For example, the corresponding salient
1070-9908/$26.00 © 2009 IEEE
Authorized licensed use limited to: Shanghai Jiao Tong University. Downloaded on November 6, 2009 at 23:03 from IEEE Xplore. Restrictions apply.
下载后可阅读完整内容，剩余3页未读，立即下载
3488 浏览量
754 浏览量
880 浏览量
2022-09-19 上传
2022-09-20 上传
209 浏览量
weixin_38720762
- 粉丝: 5
- 资源: 943
我的内容管理
展开
