International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999
42
Figure 7. Flow-chart summarising patch matching procedure.
The cost function used by Morgado and Dowman (1997) is:
T = ^\a l - a 2 \+\p i -p 2 \ + \r l - r 2 \+\c l -c 2 \+...
where T is the value of the cost function, a/ is the area of patch
i, Pi is the perimeter length, and rj and q are the length and
width of the bounding rectangle.
A specific problem associated with this function is that it was
found that the area component influenced the results more than
the other attributes. To get around this problem, the value of the
area component was halved to reduce its influence. However,
since areas and lengths are being compared, it seems more
reasonable to take the square root of the area component, so that
all the components being compared have the same
dimensionality. Furthermore, with the above function, larger
patches will always produce a larger value than smaller patches.
Therefore, to ensure this is not the case, the differences in
components have been normalised with respect to patch size.
Thus, the cost function used in this study is expressed by the
equation:
l
a \ ~ a 2
2
+
Pi -Pi
+
r i ~ r 2
+
C i C 1
«1 +a 2
Pl+ Pi
r i +r i
Cj + c 2
are shown in Table 2. It is apparent from this table that by using
a combination of the cost function, the patch separation and the
overlap, correct matches can be identified. Although the cost
function is not by itself reliable, there is a clear bimodal
distribution of the patch separations and the overlap. Using this
method, correct matches can be identified and these are shown
in Figure 8.
In order to determine the best method of segmentation, the tests
were repeated for all methods of patch extraction using the
Istres images and another pair. The results do not allow any
conclusions about the best method, only that there are large
variations with image and method. For both these images it is
interesting that the number of correct matches from all
combinations varied from 0 to 5 out of total matches varying
from 39 to 366.
The correct matches were successfully used to register the
SPOT and SAR images together.
SAR patch
index
SPOT patch
index
Cost
function
Patch
separation
(pixels)
Percentage
overlap
(x 100)
2
5204
0.463
112.429
0.194
3
0,484
0.353
0.885
7
13897
0.270
225.892
0.331
9
3417
0.550
253.794
0.312
10
2901 ¡1
0.277
4.651
0.899
12
3754
0.229
6.835
0.944
13
18389
0.265
267.599
0.199
18
26749
0.240
514.928
0.372
21
13746
0.662
209.264
0.668
22
21272
0.394
275.525
0.366
24
8585
0.259
378.102
0.401
28
18558
0.306
166.298
0.493
30
21476
0.327
198.795
0.526
32
11307
0.313
59.894
0.561
33
25114
0.129
242.566
0.561
37
21579
0.137
303.653
0.652
39
20029
0.286
395.225
0.573
43
16449
0.549
43.834
0.307
45
15793
0.287
3.995
0.887
56
1339
0.414
330.546
0.230
60
11922
0.416
382.007
0.629
61
16119
0.474
123.264
0.596
71
17966
0.215
347.164
0.514
73
13154
0.714
222.182
0.290
75
6637
0.276
384.673
0.482
76
16520
0.186
386.544
0.351
77
4572
0.382
418.060
0.471
Table 2. Matching results with correct matches highlighted.
5. EDGE MATCHING
Every patch in image 2 is matched with each patch in image 1
and the matches with the lowest cost function are retained.
There will be false and multiple matches and these are reduced
by eliminating matches with large differences in centroid
values. A further technique used to improve matching is to
repeat the operation with image 2 as the master and image 1 as
the slave. To compare shape is a further check. Patches are
compared by counting the number of pixels in common when
centroids are made coincident. Those with large overlap are
considered to be the best matches. The results of these processes
The polygon matching provides directly the centroid co
ordinates of conjugate polygons and these can be used to carry
out a transformation between the two images. This
transformation can be improved by matching the detail of the
edges around the polygons. The basic method of edge matching
using dynamic programming is described by Newton et al.
(1994). This previous work used the edges of the polygons as
extracted by the segmentation. In the new method, edges are
extracted from the raw data in the region of the polygon
boundary and then matched using dynamic programming as
before. In this way, it is ensured that reliable edges are extracted