Article
A New Image Registration Algorithm Based on
Evidential Reasoning
Zhe Zhang
1
, Deqiang Han
1,
* , Jean Dezert
2
and Yi Yang
3
1
MOE KLINNS Lab, Institute of Integrated Automation, School of Electronic and Information Engineering,
Xi’an Jiaotong University, Xi’an 710049, China; zhangzsmg@gmail.com
2
ONERA, The French Aerospace Lab, Chemin de la Hunière, F-91761 Palaiseau, France; jean.dezert@onera.fr
3
SKLSVMS, School of Aerospace, Xi’an Jiaotong University, Xi’an 710049, China; jiafeiyy@mail.xjtu.edu.cn
* Correspondence: deqhan@gmail.com; Tel.: +86-131-1911-5975
Received: 30 November 2018; Accepted: 26 February 2019; Published: 4 March 2019
Abstract:
Image registration is a crucial and fundamental problem in image processing and computer
vision, which aims to align two or more images of the same scene acquired from different views
or at different times. In image registration, since different keypoints (e.g., corners) or similarity
measures might lead to different registration results, the selection of keypoint detection algorithms
or similarity measures would bring uncertainty. These different keypoint detectors or similarity
measures have their own pros and cons and can be jointly used to expect a better registration result.
In this paper, the uncertainty caused by the selection of keypoint detector or similarity measure is
addressed using the theory of belief functions, and image information at different levels are jointly
used to achieve a more accurate image registration. Experimental results and related analyses show
that our proposed algorithm can achieve more precise image registration results compared to several
prevailing algorithms.
Keywords: image registration; evidential reasoning; belief functions; uncertainty
1. Introduction
Image registration is a fundamental problem encountered in image processing, e.g.,
image f
usion [1]
and image change detection [
2
]. It refers to the alignment of two or more
images of the same scene taken at different time, from different sensors, or from different
viewpoints. Image registration plays an increasingly important role in applications of surveillance [
3
],
remote-sensing [4] and medical imaging [5].
For a collection of images to be registered, one is chosen as the reference image and the others are
selected as sensed images. Image registration align each sensed image to the reference image by finding
the correspondence between all pixels in the image pair and estimating the spatial transformation
from the sensed image to the reference image. In this paper, we just consider the image registration
between two images, i.e., there is only one sensed image together with a given reference image.
Current image registration techniques that based on image domain can be generally divided
into two categories [
6
]: the sparse methods and dense methods. There are also some methods based
on transform domain, like Fourier-Mellin transformation method [
7
]. The transform domain based
methods are often used for image registration with similarity transformation model. In this paper,
we focus on the image domain based methods.
The sparse methods [
8
] extracts and matches salient features from the reference image and sensed
image and then estimates the spatial transformation between the two images based on these matched
features. Line features (e.g., edges) and point features (corners, line intersections and gravities of
regions) all can be used for image registration. Corner features are the mostly used features and can be
Sensors 2019, 19, 1091; doi:10.3390/s19051091 www.mdpi.com/journal/sensors
