1.1 "What's so special about the Foveola shape recognition technology?"
Foveola can deal with shape-related visual tasks in a similar way
to humans. It is therefore like having simulated human eyes interrogating
digital images at enormous speed.
Teaching the system to see new shapes is simple, transparent and fast.
Shapes can be represented in a single database which allows Foveola to
"get" the visual analogy
between an 'm' and a flying seagull, or an 'L' and a gumboot.
The actual Foveola engine is implemented as an ANSI C library, making it
both fast and easily portable across hardware platforms and operating
systems.
1.2 "I get the impression
that Foveola is just another system for Optical Character Recognition.
OCR is already a mature technology, so why do I need Foveola?"
Foveola isn't simply an OCR technology - it's a general-purpose
shape recognition engine. It can be used to recognise all
shapes, from road signs to engineering components.
1.3 "Is Foveola an end-user application?"
No, it's a tool which can be built into an existing application or
embedded in your hardware device.
1.4 "What applications areas is Foveola useful for?"
We are seeing particular interest in mobile applications such as
handwriting recognition, general image recognition software such
as multimedia search tools, and hardware-based deployments such as
interactive toys.
We expect Foveola
to have valuable applications in smart, face-sensitive cameras and other
security systems. Foveola is intended as a general shape recogniser, so we
can even envisage providing visually-impaired people with some measure of
shape recognition ability (via a different sensory modality such as hearing).
1.5 "Foveola is a so-called "neuromorphic" technology.
How is it different from ordinary Neural Networks?"
Foveola is modelled on certain aspects of the behaviour of
cells in the visual cortex of higher mammals, and
its advantages over traditional Neural Networks are considerable.
Foveola doesn't
include interconnected layers of neurons with "weights" which
require adjustment by an opaque learning process involving
many training examples.
Instead, only a single training example is required for each input
shape variant that the system is required to recognise.
Additionally, the system's current knowledge is unaffected by the
addition of new shapes; there is no danger of "overtraining".
1.6 "You talk about Foveola being modelled on how humans see,
but when I see
the world, I don't view it through some tiny window."
Actually, to some extent you do! You can only really see full detail
in an area which is about the
same as that of a fingernail held at arm's length - the "foveolar
window". (This means,
incidentally, that you can identify the face of a friend at about 30m
away; at 60m you can at least tell that it's a face and
whether it's smiling or not).
In your visual field,
the sharpness with which things can be seen falls off rapidly with
increasing radial distance from the foveolar window.
For example, when you
view a school photograph you can't spot a particular child immediately,
using only your peripheral vision. You have to scan for that child
by examining each face, one by one.
When looking at any given child, you can't even be sure whether the
neighbouring child (viewed in peripheral vision only) is the one
you're looking for.
