Another aproach is a geometric transformation of a target into the 
corresponding search segment, by using the least squares method. The 
last can minimise discrepancies in pixel intensities (5) or in their 
locations (6). Least squares method gives very: accurate results, 
provided that the terrain is smooth, there is enough signal varia- 
tion, and an approximate match has been attained. Thus, least 
squares method is feasible when high accuracy is required, image 
segments are selected carefully, and after matching by a less accu- 
rate (but more efficient) technique, e.g., by sequential search. 
5. Sequential search 
Searching techniques can be differentiated according to the struc- 
ture of image data, sequence of matching trials, and the calculation 
for a single trial. 
Data can be structured in single or multiple (hierarchical) levels, 
(vide III.5). In the latter case, search proceeds from coarse to 
fine image data, thus providing for great "pull-in" and reliability 
of matching. 
For a match (single), all possible trials can be attempted (i.e., by 
the brute force approach) or only the promising trials. The latter 
implies dynamic programming with constraints. 
À matching trial can be calculated fully or it can be interrupted 
after some initial calculation indicates a failure. 
6. Hardware 
In addition to a scauner-digitizer, a computer with standard peri- 
pherals is required. For interactive operation it is desirable to 
have a graphic CRT terminal for stereo-display. Calculations can be 
carried out serially by a general purpose computer; a dedicated 
network of processors is indispensible only for real-time operation. 
Storage requirements are modest because image data are segmented, 
and off-line operation is time-relaxed. 
7. Quality control 
Quality assessment is needed for acceptance at matching and for 
performance considerations in the subsequent stages. The criteria 
can be differentiated according to internal and external. Internal 
criteria concern the degree of similarity (i.e., maximum and "peak- 
ness" of the assessment curve), the standard error (after a least 
squares geometric transformation), and the neighbourhood (e.g. , for 
gross errors). External criteria concern the parallax data in con- 
junction with geometric conditions and check data. The latter will 
be considered further in the next section. 
Trends in image matching 
The following trends are anticipated for off-line image matching: 
Improved data structures are leading to refined matching strategies 
and vice versa. Procedures are becoming more flexible, i.e., better 
adaptable to changing circumstances. External and neighbouring data 
are being increasingly used at matching. Quality control is gaining 
in importance. 
340