enhance the reliability. 
Combining (27) and (28) forms a triplex criterion: 
p=S}|A;—B;|=min 
i=l 
AND Ÿ 045= SA; * B;=max>0., (29) 
i-1 
ml 
D,— * AA; — AB, | — min 
i1 
(d) Other combinations: 
r — max AND do —0,, —Ogg-7 min; 
n. 
05421472 SAAA, * AB, max AND "Dr = min; 
i -1 
2-2 
g,,—max AND D;—-$»1|4'A,— VYB,| =miry 
&=l 
r=max AND P;=min AND D,=min 
...... 
(e) Extension 
A multi-criterion can be constructed either by some(two or 
three) different algorithms or by a same algorithm with dif- 
ferent preprocessing of signal, i.e. smoothing, edge extrac- 
tion and different pattern of image patchs. 
4, EXPERIMENTS 
The above theoretical analysis have been verified by experi- 
ments using real image data. 
(a) Disparities betwen the criterions 
Table 1 shows real image data(e.g. grey values) sampling 
from the left photograph. The corresponding image data in 
the search area of the right photograph is listed in Table 
C. The correcte shift is T=3, e.g» the first colum of Table 
| corresponding to the fourth colum of Table 2 (parallax free 
in Y direction). The matching results using single criterions 
of correlation coefficient r, covariance Ops and sum of abso- 
lute differences f, are listed in Table 3. It has been shown 
that both the criterions r and Cas result wrong correlation 
OfT=1, but this mistake can be rejected by using the dual 
criterion of O4gand f£ or y and P , because the P reachs to 
minimum at T=3,. 
(b) Capability in detecting matching mistakes 
Table 4 shows the detail of correlation along a segment of 
- 422 -