SCALE DIFFERENCE CONSIDERATIONS IN CONJUGATE FEATURE MATCHING 
Anthony Stefanidis 
Peggy Agouris 
Department of Spatial Information Science and Engineering 
and the National Center for Geographic Information Science and Analysis 
University of Maine 
5711 Boardman Hall, Rm. 348 
Orono, ME 04469-5711 
Tel: (207) 581 2180, Fax: (207) 581 2206, e-mail: {tony, peggy} @spatial.maine.edu 
Commission III, Working Group 2 
KEY WORDS: Softcopy, Image Matching, Scale Space 
ABSTRACT 
This paper addresses the problem of matching features which have been recorded, in two spatially overlapping images at 
substantially different scales. This phenomenon may be associated with foreshortening, in which case the scale differences are 
feature- and direction-dependent, or simply with the simultaneous processing of images of different scales, in which case the 
scale variations are obviously bidirectional and global in nature. We approach this problem by employing principles of scale 
space theory, which deals with the formalization and classification of signal contents and trends by examining the behavior of 
signals in various resolutions. Coarse resolutions convey only the dominant trends of a signal (corresponding to low- 
frequency information), while in finer resolutions information details (high-frequencies) are also included. When matching 
features recorded in substantially different scales in digital imagery, we are actually attempting to establish correspondences 
among different scale representations of the same object space scene. Typical matching techniques fail or perform poorly in 
terms of accuracy in such cases, because they do not consider that beyond geometric, scale differences are also of radiometric 
nature. The methodology presented in this paper proceeds by identifying scale differences among conjugate features, 
identifying proper image pyramid levels at which matching should be performed, and only then precisely matching conjugate 
features. The analysis of the matching results permits the transformation of matching uncertainties through scale space, and 
the derivation of realistic accuracy estimates. 
1. INTRODUCTION 
Matching, the task of identifying similar features in two or 
more spatially overlapping images, is a dominant research 
issue in digital image analysis, as it is a fundamental 
operation, involved in practically all photogrammetric 
applications. Despite the great advancements made in digital 
image matching, and the numerous algorithms and strategies 
developed employing geometric and radiometric similarity 
criteria to identify conjugate features, there still exist 
problematic cases, where matching fails to produce reliable 
results. The lack of sufficient radiometric variations is a 
typical example of such a case. These problems are, to a 
certain extent, adversely affecting the role of digital image 
matching for geoinformation generation, thus delaying the 
much anticipated full automation of the mapping process. 
Among the cases where matching performs poorly, 
producing unreliable or even no results at all, is the case of 
features which have been recorded in two spatially 
overlapping images at substantially different scales. This 
phenomenon can be associated with isolated features within 
a pair of images of otherwise similar scales, or with the 
processing of images of overall different scales. The first 
8 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
case is rather object-oriented and its occurrence is dependent 
on specific image capturing and object shape combinations. 
The latter is an issue which is expected to receive much 
higher attention in the near future, as it is inherently 
associated with three-line sensor imagery (e.g. MOMS) 
which is becoming more widely available [Schneider & 
Hahn, 1992], while research also moves towards the fusion 
of aerial and satellite digital imagery for geoinformation 
extraction [Gruen et al., 1995], or the integration of digital 
imagery within geographic information systems [Agouris et 
al.; 1996], whereby digital imagery of various scales is 
combined during the performance of complex digital image 
analysis processes. 
In this paper we examine the problems occurring when 
attempting to match conjugate features whose images differ 
in scale. The presented method employs scale space 
concepts for the identification and accommodation of scale 
differences in matching. 
2. SCALE SPACE THEORY 
The information content of a signal is encoded in its values 
and their variations. These variations occur over a wide 
   
range of spati 
major signal t 
localized tren 
limited areas. 
macro- and m 
cognitive prc 
often intuition 
to be algorith 
by machine-s 
The concept 
multiple scale 
research in 
Tanimoto & | 
has been for: 
processing co! 
the papers of 
[Witkin, 19 
identification 
values of sign 
in various res 
signal defined 
(m+1)-dimens 
additional par 
Digital imag 
functions defi 
space is the 
representation 
Fx, x): com 
Uy ss) 
resolutions (te 
values (s9,5;, 
scale space f2 
scale space fa 
is decompose: 
Fig. 1: A. 
signal is at 
For different 
families of 
generation is