445
AUTOMATED SYSTEM FOR IMAGE INTEGRATION IN REMOTE
SENSING
Malcolm E. Fuller & Manfred Ehlers
Department of Surveying Engineering
University of Maine
Orono, Maine 04469-0110
E-mail: ehlers@mecanl.bitnet
ABSTRACT
In the integration of remotely sensed multisensor, multitemporal, multispectral, and
multiresolution image data, one of the more time-consuming tasks has been the rectification
of different images to a mutual geometric basis. This is due to the difficulty of automating
the process of selecting adequately distributed and identifiable control points. This paper
outlines such an automated process involving: control point identification, matching, and
correlation. Result of preliminary tests applied to the system are also reported.
KEY WORDS: image matching, Moravec operator, partitioning, correspondence,
correlation, fine tuning.
INTRODUCTION
An important step in the integration of imagery is placing
images in exact pixel by pixel registration with each other.
There are many integrative applications (e.g. modeling
global environmental cycles, multitemporal change detection,
GIS integration) which require this process. However, to
date it has remained one of the few image processing
functions which have not been automated. Image registration
still requires, to a large degree, time-consuming human input
for the selection and matching of control points (CPs) in the
images to co-register.
An automated image registration approach would make the
study of multisensor, multitemporal, multiresolution, and
multispectral image datasets much more efficient. This is
particularly true as the size of these datasets increase to
include more information, and computers become powerful
enough to handle such a large quantity of data.
Our research in the area of image registration has been
directed toward developing a prototype software package
which will select, match, and fine tune CPs with as little
human interactive input as possible (i.e. the user should only
have to enter input/output file names along with some
processing parameters). This paper will explain the approach
taken for this process, as well as a description of the
prototype software package as it currently stands.
Preliminary test results are also included.
METHODOLOGY OF AUTOMATED IMAGE
REGISTRATION
In this paper, an image which is already in a desired
geometric basis (e.g. rectified to UTM coordinates) is called
the reference image. An image which is not in a desired
geometry (e.g. raw scanned data), and must be transformed
in order to match the reference image is called the subject
image.The initial strategy taken toward the registration of a
subject image to a reference image is similar to one described
by Luhmann and Ehlers [1984] in the development of an
image matching system. The process of co-registering
images is broken up into six steps:
1 .Preprocessing of the images, which may include filtering,
scale change, window definition and other corrections to
enhance the matching process.
2.Selection of candidate control points (CCPs) in each
image to co-register.
3. Matching of the CCPs in each image, to determine which
points in the reference image correspond to those in the
subject image.
4. Fine tuning of the matched CCPs to determine the actual
location of points in the subject image to subpixel accuracy.
5. Calculation of the rectification coefficients for the
transformation from reference to subject image using the
fine-tuned control points (CPs).
6. Resampling of the subject image into the geometry of the
reference image.
In all of these steps except the first and the last, CCPs which
have been determined to be ambiguous or otherwise
unsuitable as control points are eliminated from further
consideration. At the end of the process there should be
considerably fewer CCPs than were chosen initially, but
those points should be adequate for high-precision image
registration.
CCP selection
A human interpreter normally tries to choose control points
in locations which contain easily identifiable surrounding
texture and contrast (e.g. road intersections, bridges,
building corners). Most attempts to automate the selection
process rely on functions, called interest operators, which
also locate areas in a digital image with a high surrounding
contrast. Unlike a human interpreter, interest operators do
not attempt to find CCPs which are topographically
significant [Piechel 1986]. Suitable CCPs extracted by
interest operators may be located in areas where a human
would not normally select control points, yet show enough
local contrast in the algorithm's eye for possible use.
There are several interest operators that have been developed
for use in extracting suitable points. Most of these can be
found in digital photogrammetry and computer vision
literature [Moravec 1980, Nagel 1981, Forstner 1985,
Piechel 1986, Luhmann and Altrogge 1986]. Although
originally developed for robotic vision applications, the
Moravec operator was implemented first in our software
package because it has been shown in previous research to