227 
AREA BASED MATCHING OF COLOUR IMAGES 
Michael Hahn and Claus Brenner 
Institute of Photogrammetry 
Stuttgart University 
Keplerstr. 11 
70174 Stuttgart / Germany 
phone : 0711-121-3397 
fax : 0711-121-3297 
e-mail : Michael.Hahn@ifp.uni-stuttgart.de 
KEY WORDS: Colour, multichannel image matching, quality analysis 
ABSTRACT 
Area based matching of intensity images is a well known technique applied to solve various photogrammetric 
tasks like parallax measurement, point transfer, orientation of cameras, DTM reconstruction and others. The 
intensities of two or more images are the observables of a least squares estimation process which aims at deriving 
the parameters of a geometric model. For matching two images the most widely used geometric model is an 
affine mapping between local areas of the image pair. Experimentally verified is the high precision of area 
based matching which is about 1/10 th of the pixel size. Roughly this rule of thumb holds also for the different 
generalizations of modelling the least squares matching problem including multi-image, object-space oriented, 
geometrically constrained, and other variations. 
Up to now only little attention has been given to the extension of the matching model to colour or multispectral 
images. Colour is generally considered to be an important clue for identification and recognition processes. 
'The purpose of this paper is to investigate quality differences between an area based matching of colour or 
multichannel images and images with just one channel. The formulation of multichannel image matching is 
presented by using a vector valued image function. For the experimental investigation aerial colour images are 
used. Precision of colour image matching is derived and a comparison is given to the results of intensity image 
matching. : 
1. INTRODUCTION 
Colour vision is probably the most intensively studied sensory process in human vision. The purpose of colour 
vision is to extract useful information of the spectral properties of object surfaces (Bajcsy et al., 1990). Many 
applications of colour image processing occur in biology and medicine in which, for instance, true-colour images 
are needed for the interpretation of stained microscope slides by pathologists (Ledley et al., 1990). In machine 
vision applications most interest is in image segmentation and in recognition and identification of coloured 
objects. Today colour is considered to be one of the major features in identifying objects. 
Colour image segmentation using colour edges is primarily an extension of intensity segmentation to multi- 
dimensional processing with extra information provided by colour. The first physical approaches of colour 
image segmentation have been introduced only recently (Klinker et al., 1988). In addition to changes of sur- 
face reflectance other aspects like variations due to shading, shadows and highlights are taken into account. 
Furthermore multiple reflections between objects, variation of illumination and others are studied in colour 
vision. 
Only little attention is drawn to the matching of colour images. If a colour template library of representative 
samples of known objects is given then matching may be used to solve for the recognition of those objects in 
images. The underlying principle for recognition is that “objects that look alike are likely to be alike”. Location 
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop “From Pixels to Sequences”, Zurich, March 22-24 1995