Using a smaller window (MY,MX) and minimizing this function (zero 
crossing search), the position of edges can be determined. This is 
not done here. Instead, a search is made for significant structure 
and therefore the average curvature is maximized. 
Operator : 0 2a 2b 2c 
Least squares shan SD SD R SD R SD R 
  
AD 40.Cfipixeldt :05444511157.413401714ud 7.35540. 72> 72i:20.64 
3 
.5 -0.65 1.4 -0.60 
3 
  
A1 (0.01) 2.1 1.4 -0.67 1 
A2 s{ "3 1.9 1.2 -0.68 1.3 -0.65 1.2 -0.65 
HO (grey value) 2.0 4:1 4:1 1.2 
Hi (0.01) 6.4 4,1 4.0 3.1 
SP Oper ata haat oti Ont 258 1% 488050 Bas yen: 20 apensene Eon dé 
Cross correlation : CC CC R CC R CC R 
mean maximum : 10.72 0.91 40.47 0.91. 40.52. 0.90 40.46 
Variants 2a and 2b show little difference. There is an improvement in 
accuracy of AQ over the previous method (comparison of sums) and the 
following method (comparison of products). This could be expected 
because the grey level gradients appear as factors in the adjustment. 
Determination of tilt A1 is also improved since well-structured 
features throughout the window are used, and not only edges in the 
y-direction. 
Maximizing 2nd differences (2c) gives a similar result, but the 
dependence of A0 and A1 on the interest value is not so high. 
3.3 Products 
The summation of products of neighbouring pixels can also be used to 
evaluate the selected feature. 
IV 22. A gíxl,y) *.g(x2,y) ) where x2 = xi + 1 
I +3 1 = 1..MY , jis 1..MX 
If this function is then normalized, an auto-correlation coefficient 
for shift in x-direction is obtained. The feature within the window 
can then be evaluated by this function. 
a) Maximizing the auto-correlation function corresponds to a search 
for structures with mainly low frequency components, since a small 
shift in the window (1 pixel) shows little change in the structure, 
i.e. auto-correlation remains close to unity. 
b) In contrast to maximizing, minimizing the coefficient finds 
Structures with high frequency components, since a small shift in 
the window produces a large change in auto-correlation (towards 
zero). 
- 570 -