rval 
d by 
dom 
fast 
ting 
IM. 
d in 
the 
'S as 
vhat 
n of 
zital 
the 
n its 
Fo 
that 
can 
/ be 
rom 
tour 
ubic 
red, 
ctor 
ntly 
thin 
reo 
d of 
k is 
r by 
fore 
; or 
All 
e or 
me 
ting 
The 
| be 
L In 
(ter, 
and 
tine 
em. 
and 
jhic 
are 
drawn, a seamless orthoimage, contour image and combined 
orthoimage with contours can be mosaiced automatically. 
Geometric and radiometric corrections are made along the seam 
line on the overlap regions. The geometric correction are based on 
the DTM merging. The differences between the DTMs can be 
adjusted in the overlap region by recreating that part of the over- 
riding DTM according to all coincident points of the overlapping 
DTMs so that any small geometric misalignment can be 
eliminated. The mosaic process eliminates the radiometric 
differences of the orthoimages by a weighted average algorithm. 
The geometrically continuous and radiometrically seamless 
orthoimages are then mosaiced over the entire required area. 
Any type of original images or rectified images can be mosaiced 
after interactively measuring some conjugate point pairs semi- 
automatically (in a way similar to that in the relative orientation). 
The transformation is determined by the measured points. A 
relative rectification ensures that the mosaiced image is seamless. 
Processing is performed randomly between two neighbouring 
images, step by step, to ultimately generate the mosaiced image of 
the entire block.. 
2.13 Visualisation 
Users can view the results of every stage of the photogrammetric 
process from raw images, through epipolar images in full stereo, 
orthoimage, contours, orthoimage combined with contours and the 
landscape perspective model (as shown in F igure 1). The dynamic 
landscape (in mono or stereo) is usually based on the DTM and 
orthoimage, but can be generated earlier in the process from the 
match results and an epipolar image. 
   
Figure 1. Three Gorges - China 
2.14 Output 
Raster data can be superimposed with vector data. Combining 
orthoimage, contours, graphics of objects, grid and map frame 
with necessary annotation input by the user, the image map can be 
output as hardcopy, or converted to many image formats, such as 
TIFF, IRIS RGB, SUN Raster, BMP and JPEG. The stereo 
landscape can be “grabbed” from any aspect and ground distance 
for output. The vector data, including contour data and digitised 
points, lines and polygons can be plotted for traditional vector 
maps with a frame and annotation. The output symbol library is 
the same as that in the digitise module. The DTM and vector 
contours are stored in ASCII format, and the vector data can be 
427 
  
converted to other proprietary vector formats. 
3. IMAGE MATCHING 
The most important function of a DPW is image matching for 
three dimensional data reconstruction. Previously, there have been 
some weaknesses in traditional image matching algorithms. New 
robust image matching algorithms are now being researched and 
used in a practical DPW to attain reliable results. The global 
image matching algorithms are suitable for this purpose. 
3.1 The Weaknesses of Traditional Image Matching 
Algorithms. 
Image matching, as in plate matching, is a pattern recognition 
problem. Grid points on the left image are samples of the objects, 
and the points on the right image are their classifications. The 
image match determines which sample belongs to which 
classification. In traditional image matching algorithms, some 
criterion, such as the maximum correlation coefficient, is used to 
decide that a sample is or is not to belong to a certain 
classification. Firstly, they do not take spatial relationships into 
account, and further, they do not use the matching results in the 
neighbourhood to adjust the global results of the match. Secondly, 
it is virtually impossible for the probability that the error 
classification is zero for any criterion of classification, and 
therefore wrong results for the image match are unavoidable. The 
results from traditional image matching algorithms are therefore 
inharmonious and unreliable. 
3.2 Bridge Mode of Image Matching 
The most important difference between images in a stereo pair is 
the effect of geometric distortion by ground slope. Therefore, the 
sizes of conjugate image windows should usually be different. 
The method of creating the image windows in a rectangular form 
centring the target on the matched point did not consider this 
distortion. This has been classically referred to as the Centre 
Mode of image matching. The image windows in the Bridge 
Mode (Zhang Z., 1989 and 1990) are created between two target 
points and their candidate conjugate points separately. For 
example, i and k are two points in the same epipolar line of the 
left image, and j and 1 are their candidate conjugate points in the 
conjugate epipolar line of the right image. The image segment 
(i, k) is defined as a target window and the image segment 
(j, 1) is defined as search window. The size of the search 
window (j, 1) is different to the size of the target window (i, 
k),thusaresampling for (j, 1) relativeto (i, k) should be 
completed to ensure their same size before comparing their 
similarity. The distortion caused by ground slope will be rectified, 
and the quality of the image match will be improved. In fact, with 
the concept of bridge mode, each interesting point being matching 
can be related with its neighbourhood to extend a single point 
matching into global image matcing. 
3.3 Probabilistic Relaxation of Global Image Matching 
Probabilistic relaxation is an effective method used frequently in 
image segment, edge extraction, analysis of light flow and pattern 
recognition. Relaxation processing should be a useful technique 
for using contextual information to reduce local ambiguity and 
achieve global consistency in the global image matching problem. 
It is basically a parallel execution algorithm. Being applied in 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996