I
maging
& M
achine
V
ision
emerging intelligent vision systems. By build-
ing on established designs, decades of research
and evolutionary artificial intelligence, HSAs
provide the ideal platform for the next gener-
ation of machine vision systems.
The complexity involved with understand-
ing the world through images should not be
underestimated. The vision systems found in
nature have evolved over millions of years,
yet their digital counterparts have only been
in development for mere decades. Neverthe-
less, software running on advanced process-
ing platforms can now be seen as comparable
in its ability to tackle this complex challenge.
Neural networks have been in use for many
years, but recent leaps in processing power
mean their use is no longer compromised by
the platform’s ability to match their poten-
tial, enabling the adoption and development
of even bigger and, more importantly, much
deeper (more layered) neural networks.
Indeed, while limitations in processing power
meant neural networks may have needed to
be simplified in the past, modern SoCs are
more than able to support highly complex
networks with many layers.
Other forms of artificial intelligence are also
seeing the benefits of more powerful process-
ing architectures. A leader in this field is x86
architecture, which has always been at the
forefront of adopting new technologies. It
successfully combines instructions optimized
for streaming and vector operations, devel-
oped over many years, with new technolo-
gies such as Shared Virtual Memory. All of
these innovations are employed in the latest
heterogeneous system architectures, which
allow software engineers to make full use
Figure 1. Example of what the IVS70 stereo camera sees with standard office corridor lighting
and non-optimized decoding at 5 Mpix.
