ISPRS Commission III, Vol.34, Part 3A „Photogrammetric Computer Vision‘, Graz, 2002
AUTOMATED IMAGE REGISTRATION USING GEOMETRICALLY INVARIANT
PARAMETER SPACE CLUSTERING (GIPSC)
Gamal Seedahmed and Lou Martucci
Remote Sensing and Electro-Optics Group
Engineering Physics Division
Pacific Northwest National Laboratory”
902 Battelle Blvd. Richland, WA, 99352 USA
Gamal.Seedahmed@pnl.gov
Lou.Martucci@pnl.gov
Commission III, WG III/I
KEY WORDS: Automation, Image Registration, Hough Transform, Geometric Invariance, Clustering
ABSTRACT:
Accurate, robust, and automatic image registration is a critical task in many typical applications that employ multi-sensor and/or
multi-date imagery information. In this paper we present a new approach to automatic image registration, which obviates the need
for feature matching and solves for the registration parameters in a Hough-like approach. The basic idea underpinning GIPSC
methodology is to pair each data element belonging to two overlapping images, with all other data in each image, through a
mathematical transformation. The results of pairing are encoded and exploited in histogram-like arrays as clusters of votes.
Geometrically invariant features are adopted in this approach to reduce the computational complexity generated by the high
dimensionality of the mathematical transformation. In this way, the problem of image registration is characterized, not by spatial or
radiometric properties, but by the mathematical transformation that describes the geometrical relationship between the two images or
more. While this approach does not require feature matching, it does permit recovery of matched features (e.g., points) as a useful
by-product. The developed methodology incorporates uncertainty modeling using a least squares solution. Successful and promising
experimental results of multi-date automatic image registration are reported in this paper.
1. INTRODUCTION
The goal of image registration is to geometrically align two
or more images so that respective pixels or their derivatives
(edges, corner points, etc) representing the same underlying
structure (object space) may be integrated or fused. In some
applications image registration is the final goal (interactive
remote sensing, medical imaging, etc) and in others it is a
required link to accomplish high-level tasks (multi-sensors
fusion, surface reconstruction, etc). In a multi-sensor context,
registration is a critical starting point to combine multiple
attributes and evidence from multiple sensors. In turn, multi-
sensors registration or fusion can be used to assess the
meaning of the entire scene at the highest level of abstraction
and/or to characterize individual items, events (e.g. motion),
and other types of data.
The sequential steps of feature extraction, feature matching,
and geometric transformation have evolved into a general
paradigm for automatic image registration, (see Brown,
1992). Many algorithms have been invented around this
paradigm to handle the automatic image registration with a
major focus on solving the matching (correspondence)
problem. The basic idea behind most of these algorithms is
to match image features according to their radiometric or
geometric properties using a pre-specified cost function to
assess the quality of the match; (see Dare and Dowman,
2001; Thepaut et al., 2000; Hsieh et al., 1997; Li et al., 1995;
Wolfson, 1990). While these methods have certain
advantages in computing the transformation parameters in a
single step and in retaining the traditional way of thinking
about registration in the sense of identifying similar features
first and then computing the parameters of the geometric
transformation, they have considerable drawbacks in meeting
the current challenges of image registration. First of all, they
require feature matching, which is difficult to achieve in a
* Operated by Battelle for the US Dept. of Energy (DOE)
A - 318
multi-sensor context since the common information, which is
the basis of registration, may manifest itself in a very
different way in each image. This is because different sensors
record different phenomena in the object scene. For instance,
take a radar image vs. optical image. Second, the feature
extraction algorithms are far inferior in the sense of detecting
complete image features. For instance, missing information
such as edge gaps, and occlusion are two famous examples
that could lead to incorrect matching
In the late nineties and through 2001, Hough Transform
(HT)-like approaches emerged as a powerful class of
registration methods for image and non-image data. This new
class of methods provides a remedy to the above-mentioned
problems and considers different strategies to reduce the
computational complexity that hampered the wide use of the
original HT (Hough, 1962). In comparison to the previous
approaches, this new class is a correspondence-less strategy
since it does not use feature correspondence to recover the
transformation. Instead, a search is conducted in the space of
possible transformations. The Modified Iterative Hough
Transform (MIHT) is a representative method that belongs to
Hough-like approaches. MIHT is developed to solve
automatically for different tasks such as single photo-
resection, relative orientation, and surface matching, (see
Habib and Schenk, 1999; Habib et al, 2000; Habib and
Kelly, 2001*^: Habib et al., 2001). In MIHT, an ordered
sequential recovery of the registration parameters is adopted
as a strategy to reduce the computational complexity. This
ordered sequential solution considers quasi- invariant
parameters to reduce the computational complexity. These
parameters are associated either with specific locations that
de-correlate them or with the selection of data elements that
contribute to specific parameter(s). The basic idea behind
the Hough-like approaches, such as MIHT, is the exploitation
of the duality between the observation space and the