149
STUDY ON TIE POINT SELECTION FOR CO-REGISTRATION OF DIFFERENT
RESOLUTION IMAGERY
Zhen Xiong, Yun Zhang
Department of Geodesy & Geomatics Engineering, University of New Brunswick,
15 Dineen Drive, PO Box 4400, Fredericton, NB, Canada E3B 5A3
v009v@unb.ca, yunzhang@unb.ca
Commission VI, WG VI/4
KEY WORDS: Co-Registration, Different Resolution Imagery
ABSTRACT: Image registration is a hot research topic in the field of remote sensing. Basically image registration includes three
steps, namely tie point selection, mapping function parameters refinement, and image resampling. Most research emphasis focuses
on how to extract tie points automatically and how to reduce image distortion caused by relief variation. No matter manually tie
point selection or automatically tie point extraction through algorithms, such as Harris and wavelet et al, the most commonly tie
points extracted by the automatic algorithms are comers, junctions, high curvature gradient and line ends. These feature points are
the point where the gray gradient changes steeply. Therefore, comer points, such as building comers and road intersection comers,
are usually extracted as tie points by these algorithms. This kind of tie points can work well for the co-registration of images with
the same resolution. But for different resolution images, because each image has different pixel size, there is a relative sampling
error between images caused by different pixel size. Theoretically this kind of tie points (comer points) is not truly tie points,
because they are not truly conjugate points. Therefore the comer points are not suitable for co-registration of different resolution
images. In this paper, we proposed a new kind of tie points in stead of comer points—gravity center points. Compared with the
comer points, the gravity center point is the central position of symmetric objects. Although central points are seldom extracted by
the gradient based algorithms as tie points, it is really suitable for registration of different resolution images. In this paper, the
accuracy analysis of comer and gravity center is presented. Experiments show that the central position has obvious accuracy
advantage to the comer position. Finally discussion and conclusion are provided.
1. INTRODUCTION
Image registration needs a lot of tie points. But manual tie point
selection is really a boring and time consuming work. Therefore,
a large amount of automatic interest point extraction algorithms
are developed. Tie points can be extracted in space domain and
frequency domain of an image. Comers, junctions, high
curvature gradient, gravity center, and line ends all are the
interest points. “A wide variety of interest point and comer
detectors of space domain exist in the literature. They can be
categorized into three classes: contour based, intensity based
and parametric model based methods. Contour based methods
first extract contours, and then search for maximal curvature or
inflection points along the contour chains, or do some polygonal
approximation and then search for intersection points. Intensity
based methods compute a measure that indicates the presence of
an interest point directly from greyvalues. Parametric model
methods fit a parametric intensity model to the signal. They
often provide sub-pixel accuracy, but are limited to specific
types of interest points, e.g., L-comers” (Cordelia, et al., 2000).
Fourier transform and wavelet transform are widely used for
feature extraction in frequency domain. Gang (2004) developed
a wavelet-based feature extraction technique and relaxation-
based image matching technique to find tie points for image
registration. Chen et al. also use wavelet transformation to
detect comers (1995). The algorithms of extracting feature
points in frequency domain are essentially gradient based and
usually can extract comer features.
“The parametric model used by Rohr (1992) is an analytic
junction model convolved with a Gaussian. The parameters of
the model are adjusted by a minimization method, such that the
template is closest to the observed signal. In the case of a L-
comer the parameters of the model are the angle of the L-comer,
the angle between the symmetry axis of the L-comer and the x-
axis, the greyvalues, the position of the point and the amount of
blur. Positions obtained by this method are very precise.
However, the quality of the approximation depends on the
initial position estimation. Rohr uses an interest point detector
which maximizes det(A) (equation(2)) as well as the
intersection of line segments to determine the initial values for
the model parameters. Deriche and Blaszka (1993) develop an
acceleration of Rohr’s method. They substitute an exponential
for the Gaussian smoothing function. They also show that, to
assure convergence, the image region has to be quite large. In
clustered images the region is likely to contain several signals,
which make convergence difficult. Baker et al (1998) propose
an algorithm that automatically constructs a detector for an
arbitrary parametric feature. Each feature is represented as a
densely sampled parametric manifold in a low dimensional
subspace. A feature is detected, if the projection of the
surrounding intensity values in the subspace lies sufficiently
close to the feature manifold. Furthermore, during detection the
parameters of detected features are recovered using the closest
