m 
R(u,n) - S V = Vk [Wi Si (8) 
k-l 
Where 
u : Feature index 
n : Image index 
m : Number of constraints 
Vu : Value of the kth constraint of feature U, 
W, : Normalisation coefficient 
S, : Strength of a constraint 
This rating function can be converted to range [0,1] by: 
a 
fme LE (9) 
a * R(u,n) 
in which a is a constant that determines the change rate. 
4. EXPERIMENTAL RESULTS 
The proposed procedure was applied to several wounds. A 
sample image of an artificial wound is shown in Fig. 5. 
  
edi (b) 
Figure 5 
a) Original Image 1 
b) Modified Image by Homomorphic Filter 
The total number of extracted points was 790, 747, and 759 
points for first, second, and third image, respectively, with an 
average number of 765 and standard deviation of 22. The mean 
point size was 23 pixels. Being less dependent on intensity 
values, this technique for point detection looks more robust in 
wound measurement procedure in comparison to Forstner 
operator. Using Forstner operator, an average number of 854 
points was detected, but with a standard deviation of 93. The 
extracted points of the first image is shown in Fig. 6. 
  
Figure 6. Binarized Image and Detected Points 
With 14 micron epipolar band width, 7 micron intersection 
uncertainty for epipolar lines, 100 mm depth range estimate, 
and 5 matching constraints, the initial correspondence list 
included 712 points with 253 ambiguous matches. Applying the 
local matching constraints led to 67% reduction in the number 
of mismatches. By the aid of global matching constraint, 95% 
of the 83 remaining ambiguous points were rejected. Finally, a 
cloud of 459 robust 3D points was reached. The resulting 
reconstructed surface is depicted in Fig. 7. 
   
Ei 
Figure 7. Reconstructed Wound Surface 
Having a set of reliable 3D points on the wound and 
surrounding surfaces, the desired final medical products such as 
area, volume, and direction of wounds can be easily found by 
numerical computations. The results show considerable 
reduction in the number of remaining ambiguities. 
S. FUTURE WORK 
Future research will aim at finding the optimum pattern for this 
specific problem. The attributes of the dot pattern can be 
determined as follows. 
Shape. Since perfect camera calibration is impossible to 
achieve, a more flexible symmetric interpretation of the 
epipolar constraint is more appropriate (Ozanian, 1995) : A pair 
of features (S, S^) satisfies the epipolar constraint if and only if 
the epipolar plane passing through the centroid of S intersects 
S". We have to look for uniquely occurring subpatterns. 
Intensity. To make the subsets of the pattern as non-repetitive 
as possible, the pattern dots can be composed in different 
intensity levels, all above the background upper limit and with 
a jump in between the levels significantly higher than the 
expected noise. 
Density. A natural question to ask is how should the knowledge 
of the uncertainty of the points of interest and the Fundamental 
matrix influence the search of the corresponding point. The 
search should take place in an area of the Target Image, called 
the epipolar band, that is delimited by the two branches of a 
hyperbola (Faugeras, 2001). The epipolar band is the set of line 
in the Target Image corresponding to the points in the Source 
Image that fall inside the ellipse of uncertainty. Assuming a 
satisfactory calibration procedure, we neglect the discrepancies 
from the straight line (Fig. 1). Unsolvable correspondence 
situations occur when pattern dots appear very close in the 
images such that the distance between dots is less than the 
search band radius. 
Size. The functional relation between object details and their 
reproduction in an image is given by Shannon theory (Azizi, 
1999) : Only those object structures that are at least twice as 
large as the sampling interval can be unambiguously 
reconstructed from the images. Taking into account the 
projector intrinsic parameters, the optimum size of the pattern 
dots are computed. 
Another topic for future research is the use of a calibrated 
projector. The intrinsic parameters of a projector consist of 
focal length, principal point, lens distortions, magnification, 
modulation transfer function, depth of focus. Projector 
calibration can provide this information as well as the relative 
position of the projector with respect to the cameras. This 
—268— 
inf 
rob 
reg 
rec 
The 
tha 
the 
Dis 
use 
me 
cor 
cal 
apr 
anc 
tha 
Cor 
Alv 
Dis 
Vis 
Phe 
Bo 
Cai 
Net 
Fat 
Jah 
API