16 
7 8 9 
10 
12 
  
  
  
  
  
Fig. 6 Hegions merged after completion of Stage Ila. 
The removed boundaries are represented by dotted lines. Re- 
gion 1 and 2 are merged forming region 14, and region 3, 
4, 5 and 11 are merged forming region 16. 
Region 10 is non-significant and must be merged. The 
neighbour yielding the lowest merging t-score is 12. So, 10 
is merged with 12, forming 18. The final result of Stage I 
and Stage II is shown in fig. 7. 
  
14 
  
  
  
  
17 
  
  
  
  
  
  
18 
  
  
  
  
  
Fig. 7 Final result of our entire edge based region grow- 
ing scheme. In Stage IIb region 6 is merged, due to its small 
size, with region 16, the most similar neighbour, forming re- 
gion 17. In Stage Ilc region 10, which is due to a mized 
pizels boundary, is found insignificant and merged with re- 
gion 12, forming region 18. 
6 Experimental Results 
We have implemented the procedure within DIGIS, our in- 
door developed image processing software package, presently 
running on a SUN 3/60 workstation. The programs are 
written in FORTRAN-77, and consist of two independent 
modules corresponding with the prediction stage (Stage I) 
and the merging stage (Stage II). 
798 
6.1 The Test Images 
We have tested our method on a number of artificial and 
real images. We use two artificial images, see fig. 8 and 
fig. 9. The contrast Ag between background g, and object 
go is uniformly set to 100; g, = 75 and g, = 175, for both 
SB and PV1. Notice that the border pixels of PV1 are re- 
ally mixed pixels, in contrast with common synthetic test 
images, which makes our test images much more realistic. 
The images are contaminated by a zero-mean Gaussian dis- 
tributed pseudo-random noise field generated by computer, 
with o,, = 10,20, and 50, resulting in a signal-to-noise ratio 
SNR = Ag*/02 = 100,25, and 4, respectively. We use two 
real images, see fig. 10 and fig. 11. 
To reduce the noise, the image may be preprocessed by 
several types of smoothing filters (see section 3). Each filter 
has window size 3 x 3 and is applied non-iteratively. The 
threshold of the conditional average filter is set to 30. 
6.2 Artificial Images 
SB images To demonstrate the entire experimentation, 
the SB image is treated in length. Fig. 8a gives the ideal SB 
image. Fig. 8b is the image after adding a o, = 50 noise 
field, and fig. 8c shows the result of the 3 x 3 extended 
Kuwahara filter on fig. 8b. This is the input image for 
Stage I. 
  
  
  
  
  
f 
Fig. 8 SB image, size 64?, a) ideal image; b) ideal 
image corrupted with a o, — 50 noise field; c) result of 
eztended Kuwahara smoothing on b; d) result of Stage I, 
the prediction stage; e) final result of Stage II, the merging 
stage; f) outline of the regions. 
tic 
99 
th: 
ea 
sm 
sin 
are 
reg 
dis 
20 
alr 
sm 
Stz 
the