160
AN OPTIMIZATION HIGH-PRECISION REGISTRATION METHOD OF MULTI
SOURCE REMOTE SENSING IMAGES
LIN Yi a , JIAN Jianfeng b , ZHANG Shaoming 3 , XIE Feng 3
3 Department of Surveying and Geoinformatics, Tongji Univ., Shanghai 200092, China
b Ministry of Edu. Key Lab. of Computer Network and Information Security, Xidian Univ., Xi’an 710071, China
c Xi’an Research Institute of Mapping and Surveying, Xi’an 710054, China
Tel: 021-65986960, Fax: 021-65985811, Email:linyi@mail.tongji.edu.cn
KEY WORDS: remote sensing images, registration, feature point, variant moment, relaxation matching
ABSTRACT:
An optimization method that registering the remote sensing images which are from different sensors, with different resolution and
taken at different time, is presented in this paper. First the feature points are quickly extracted with the operators of gradient and
Forstner, and evenly distribution by the use of the grid technique based on the entropy. With the help of some exact control points,
Based on the matching rule of the variant moment similarity measurement and the matching strategy of the global relaxation, the
conjugated points are gotten by quickly registering the remote sensing images. Finally, the error points are eliminated by using the
quadratic polynomial model. Experiment results show, the method, with the quickly registering speed and the high accurate and
evenly distributing conjugated points, can meet the need of the image fusion and quickly updating of the remote sensing images.
1. INTRODUCTION
With the development of modem remote sensing technology,
remote sensing is growing up to multi-sensor , multi
resolution, multi-spectrum(ultra-spectrum) and multi-temporal
information acquisition and fast intellectualized processing. The
amount of data of remote sensing is rapid rise, how to dig out
the potential of mass data and to improve the efficient of dada
application is a new task on the technology development of
remote sensing image processing. Earth observation satellite
has provided more and more multi-space resolution, multi
temporal and multi-spectrum image in the same place.
Moreover, it has provided abundant data for hypsometry, map
updating, the classification of land resource utilization, crops
classification and forest classification , flood disaster
monitoring, variety monitoring in ecological resource, etc..
During the transition from the applied analysis of single sensor
image to the analysis and application of multi-spectrum, multi
sensor , multi-platform , multi-temporal , multi-resolution
image, automatic registration and merging of multi-source
remote sensing information is very important to realize the
spatial data variety detection and image date updating,
moreover, spatial registration of multi-source image the very
important step of the next image merging, its error direct
influence the result validity of multi-source image merging.
The traditional image registration method is to manual seek the
homologous between the undetermined registration image, it is
very time-consuming and arduous, and could not adapt the
demand of mass of data processing, furthermore, it has a great
subjective influence over the precision of registration.
Therefore, many researchers go out their way to look for the
automatic or semi-automatic image registration. Strunz and
other scientists have put forward a schema of automatic seeking
control point for image geometry rectification; Diamdji has
proposed a automatic registration method of different resolution
image based on wavelet transform. The both proposals take the
low-resolution image as reference image and lose the high
resolution information. Professor Zhang Zuxun has put forward
a full automatic remote sensing image registration method
based on probabilistic relaxation total matching of multistage
image, it could use in different temporal and resolution image
registration, but this method didn’t consider the uniform
distribution of matching points and how to reject the
mismatching points; Zhang Jixian has proposed a fast automatic
image registration method by using polynomial model to total
rectification , feature extraction combined with pyramid
template matching. But the multilayer template matching just
used some local information around the undetermined
registering point, and it couldn’t use the strong correlation
between undetermined registering point and registered point,
moreover, it didn’t consider the global consistency of result,
and it is lack of registering reliability.
In this paper, a registration method of multi-resource remote
sensing image is presented. The basic thought is to fast extract
the feature points with the operators of Roberts gradient and
Forstner, at mean time, to control the evenly distribute by using
of the grid technique based on the entropy. With the initial
steering by a few of exact control points, the homologous are
gotten by quickly registering the remote sensing images, based
on the matching rule of the variant moments similarity
measurement and the matching strategy of the global relaxation.
Finally, the error points are eliminated by using the quadratic
polynomial model. Experiment results show that the method,
with the quickly registering speed and the high accurate, can get
the evenly distributing homologous and meet the need of the
quickly updating the remote sensing images.
2. FEATURE POINTS EXTRACTION
There are some ordinary operator for feature points extraction
such as Moravec one and Forstner one[14], etc.. Moravec
operator is more convenient than Forstner one, however
Forstner operator has better precision and could indicate the
type of feature points. In this paper, we prefer Forstner one
because of its better precision. Before extracting the feature
