Hype & Spin in Signal &

Image Processing Research

Fads & Fallacies filed by
Oscar Ducat



Fads
Leonardo's Lady with Ermine
The Single Pixel Camera

Does random sampling
(compressive sensing)
achieve better results
than regular under-sampling?
Structural Similarity (SSIM) 

 Is the index for the similarity
of two images an improvement 
on the noise visibility function?

The citation juggernaut spins
out of control and few researchers
dare suggest that a much simpler
formulation just might be as good,
if not better...
Leonardo's other
              lady

Certainly no convincing evidence in this experiment.
Compare with simple under-sampling of images.
SSIM Simplified


The citation juggernaut plows on regardless
Google Scholar tracking...
1307 and counting...






Fallacies: Incites & Insights
Shannon entropy

defined by the intensity histogram
 measures image complexity

Infinite width of SINC function

rules out its use in accurate interpolation
of discretely sampled band-limited
images

Robust digital watermarks

immune to geometric transformations
can only be embedded in
the transform domains
(i.e. Fourier, DCT, Mellin, and wavelet)
The human ear is insensitive 

to the phase of audio signals
Pixel value histogram gives no indication
whatsoever of the spatial complexity
or compressibility of a
2-D (or even 1-D)
signal
Fourier based sinc interpolation is easy
to implement and can actually
achieve lossless
resampling
Dirichlet sinc is periodic
Direct spatial domain embedding is
generally easier, faster and
not impossible to
comprehend
The evidence for phase sensitivity
is piling up.  Digital (DSP) filters
now allow perfect phase and
amplitude correction
in HiFis
Insights

E.T. Jaynes, 1965

It is clearly meaningless to ask
“What is the entropy of the [system] crystal?"
unless we first specify the set of parameters
which define its thermodynamic state.

E.P. Wigner
“Entropy is an anthropomorphic concept"

K.G. Larkin, 2016
Make Shannon Information symmetric in 1D
and everything else
follows naturally
into 2D and beyond.  Check preprint on arXiv

Insights

Insights
An example of affine invariant functions
embedded directly in spatial domain (LHS):

Affine distortions and corresponding detection
                    peaks
Correlation detection survives all affine
distortions (
RHS is polar spatial domain ).
Note flashing bright lines are (negligible)
image edge effects.

Insights



Pure Hype & Spin
"That was merely the
algorithms panicking...

...you know, in a metaphorical sense.
And just, you know, they were
freaking out because what
happened in China was not something
that was in anybody's... math."


Placeholder TBA Placeholder TBA
Who pretends to understand the interplay of these
                algorithms?
another early entropy evaluation
don't be so negative

Still under construction in late 2016. Last updated on 18 November 2016