Article
Robust Non-Rigid Feature Matching for Image
Registration Using Geometry Preserving
Hao Zhu
1,
* , Ke Zou
1
, Yongfu Li
1
, Ming Cen
1
and Lyudmila Mihaylova
2
1
Key Laboratory of Intelligent Air-Ground Cooperative Control for Universities in Chongqing,
and Automotive Electronics and Embedded System Engineering Research Center, College of Automation,
Chongqing University of Posts and Telecommunications, Chongqing 400065, China;
kezou18@163.com (K.Z.); liyongfu@cqupt.edu.cn (Y.L.); cenming@cqupt.edu.cn (M.C.)
2
Department of Automatic Control and Systems Engineering, University of Sheffield, Mappin Street,
Sheffield S1 3JD, UK; L.S.Mihaylova@sheffield.ac.uk
* Correspondence: zhuhao@cqupt.edu.cn
Received: 24 April 2019; Accepted: 14 June 2019; Published: 18 June 2019
Abstract:
In this paper, a robust non-rigid feature matching approach for image registration with
geometry constraints is proposed. The non-rigid feature matching approach is formulated as a
maximum likelihood (ML) estimation problem. The feature points of one image are represented
by Gaussian mixture model (GMM) centroids, and are fitted to the feature points of the other
image by moving coherently to encode the global structure. To preserve the local geometry of these
feature points, two local structure descriptors of the connectivity matrix and Laplacian coordinate
are constructed. The expectation maximization (EM) algorithm is applied to solve this ML problem.
Experimental results demonstrate that the proposed approach has better performance than current
state-of-the-art methods.
Keywords:
image registration; non-rigid feature matching; local structure descriptor;
Gaussian mixture model
1. Introduction
Image registration is a fundamental task in many fields, such as computer vision, robotics, medical
image processing, and remote sensing [
1
–
4
]. The main purpose of image registration is to align two
or more images of the same scene taken from different viewpoints, at different times, and/or by
different sensors.
Many algorithms have been developed for image registration. It can be roughly divided into
area-based and feature-based methods. Area-based methods match image intensity values directly.
They mainly include the cross-correlation (CC) methods, the Fourier methods, and mutual information
(MI) methods. The normalized CC method is a classic in the area-based methods [
5
]. The similarity of
window pairs from two images are computed and the maximum is considered as a correspondence.
A method based on wavelet decomposition and correlation is proposed for image registration [
6
].
The Fourier methods find the Fourier representation of the images and a subpixel phase correlation
with Gaussian mixture model (GMM) is used to register the images [
7
]. The MI method provides a
measure of dependence between two images, and a deterministic explanation for MI-based image
registration is proposed in [8].
Feature-based methods extract the salient structures, i.e., features, from the images. The extracted
features are called control points [
9
], and the feature-based methods are considered as point set
registration. The traditional for feature-based image registration approach uses a two-step strategy.
In the first step, the distances of feature points from local descriptors, such as scale-invariant Fourier
Sensors 2019, 19, 2729; doi:10.3390/s19122729 www.mdpi.com/journal/sensors
