Helmhotz's Treatise on Physiological Optics

Helmhotz's treatise discusses relationships among sensations,
perceptions, and ideas.  Several examples of visual and other sensory
illusions are cited.  He points out that humans are not usually aware
of the actual sensations but only those which enable us to construct
an idea of the external environment.  Only by careful attention and
training can some people observe sensory inputs not noticed before,
for example, the blind spot, double vision, mouches volantes.  He
mentions two opposing theories of vision: the empirical theory, that
vision comes from experience, and the intuition theory, that vision is
basically innate from birth.  It is difficult to tell how much of each
is in the actual visual system, for memory of experiences inevitably
taints the current visual perceptions of objects.  Certain illusions
do not vanish merely from understanding their causes.  People
interpret their sensory inputs in the way normally encountered, even
when they realize that they are due to a different cause.

Helmhotz, as discussed in class, also intimates the idea of a
model-based vision system.  Objects in the environment can only be
detected through sensory perceptions.  THe only characteristics we
know of the object are those which we can detect.  The perceived
object and its representation as an idea in our minds are only a model
of the object.  In this sense, there is no true nature of the object.
If there is a characteristic not detectable by us in any means, does
it exist?  Probably not until we find a means.


From the wealth of information presented in the biological readings, I
can remember a few interesting facts which may be relevant to the
development of machine vision.  Even at the sensory level, neighboring
cones exhibit antagonistic behavior, mediated by interneurons, which
thus help the perception of edges.  This is very much low level
processing and perhaps can be emulated by linked parallel processors,
each processor in charge of a "cone" or a pixel

In addition, the layout of the visual neurons in the brain is
interesting.  Nerves corresponding to the same portion of the visual
field but in different eyes are sent to the same side of the brain.
Neurons corresponding to a connected region are usually close to each
other.  (Question:  What happens to the nerves which are close to the
edges of different hemiretinas?)  It is also interesting that whereas
certain neurons have on-center and off-center receptive fields to
detect light contrasts, simple cells in the primary visual cortex
respond to straight bars at different angles, which process may help
the detection of straight lines and edges.  It would be fascinating to
emulate and understand the flow of information in human vision with a
machine given this knowledge of the biological architecture.


One problem underlying machine vision is probably the integration of
low level and high level processing.  Since visual problems tend to be
grossly underconstrained, high level models and heuristics are
desirable.  Higher level processing tend to involve more matching,
symbolic manipulations, etc.  Lower level processing contains vast
amounts of computations and transformations of images into other
images to enhance edges, segment regions, etc.  Many algorithms are
known, and they seem to me more fundamental than some of the
high level rule-based heuristics.

Higher level interpretations seem to warrant much research.  There
doesn't seem to me a consistent way of interpreting images.  One
approach heads towards line drawings and the blocks world in ideal
situations.  Another approach is the shape from shading or texture
methods where the objects in the scene do not have to be known
beforehand.


