pport a
Ver is
ics in
eye is
igure),
tential
> image
lor in
reefold
1j). On
, false
casting
.h more
lor CRT
ch may
Color
lution
Ling in
1Ised in
photo-
photo-
ligital
ced by
le and
aerial
on and
th the
tes.
essed,
de and
e will
ta and
in the
ds. À
> data
aerial
multi
in one
e of a
of a
5.
4. DIGITAL VIDEO RATE PROCESSING
The issue in electronic image processing and analysis is speed. This is
so whether it is to process large images of several thousand pixels square
in a reasonable time, or TV size images in real-time. Only a few processing
algorithms are "single point" (one arithmetic operation per pixel); ones
such as addition and subtraction, or code conversion as with look-up table
transformation. More interesting and powerful ones such as neighborhood,
transform-domain, statistical, and morphological operations require several,
tens, possibly hundreds, of arithmetic operations per pixel. The
traditional Von Newmann general purpose computers are woefully inadequate
for the demands modern image processing impose, although they can perform as
control hosts in digital image systems. The "minisupercomputer" promises
help at the high end of the power spectrum for photogrammetric image
processing while special processors (different configurations of array and
parallel processors) are also emerging with a range of capabilities.
Progress in parallel systems, Hung, 1984, Lerner, 1985, Meng, 1986, becomes
more essential when it is realized that the speed increase provided by
improved hardware technology is only running at about a factor of two every
four years, Manuel, 1986. Confounding the progress in parallel architecture
however is a relatively sluggish advance in parallel software, Wolfe, 1986.
Image processing algorithms which are repetitive present little problem, but
related applications involving artificial intelligence are difficult to
decompose and are very application dependent, Charniak, 1985, Wolfe, 1986.
4.1 Towards the Cost Effective Supercomputer
It is difficult to believe ten years have passed since Seymour Cray
ushered in the era of the supercomputer. A new generation of minisuper-
computers is emerging to fill the large gap between the Cray level of super-
computer and the superminicomputer. Formerly restricted to such
applications as weather prediction and weapons research, small-scale
engineering projects are expected to benefit from the coming
minisupercomputer boom. Photogrammetric image processing, particularly of
large data bases, can evolve by planning to integrate powerful computers
into future systems. As Table 1 shows, a race is developing to build
smaller, low cost supercomputers.
Table 1. Examples of High Performance Computers, (Ohr ‚1 1986)
Computer MFLOP's* |Price (approx.) Designation
Cray X/MP 700 $10Million Supercomputer 1
Cray-1 300 5M Supercomputer 1
Intel iPSC-VX/d6 424 0.85M Supercomputer
FPS-T-10 128 0.5M Supercomputer
FPS-T40000 (262,000) (200 M) (Proposed)
Alliant FXB 90 1M Minisupercomputer 1
IBM 3090 15 3M Minisupercomputer 1
FPS-5000 60 0.15M Minisupercomputer 1
DEC 8800 3 0.6M Superminicomputer 1
DEC VAX 8650 1.5 0.7M Superminicomputer 1
*Millions of floating point operations per second, (64 bits).
91