Truncation error feedback
has been employed successfully in the digital
coding of audio signals for about 20
years. The fundamental idea is to
gain control over spectral shape
of the truncation error noise which is
defined as the difference between the
original signal and its truncated,
rather, quantized version. The
motivation can be both perceptual and
quantitative; that is, in some
circumstances it is advantageous to have
colored truncation noise because of a
certain perceptual insensitivity,
while in other circumstances the MSE,
in what is considered to be the
spectral baseband, can actually be
reduced by coloration.
We view truncation noise and
quantization noise in basically the same way,
arising from similar phenomena.
In our report, we demonstrate and
conclude that truncation error
feedback is not successful in the
quantization of still
images, and we hypothesize why:
Hypothesis
In the images we examine,
we do not see a preponderance of quantization
error energy localized in
any particular spatial frequency region. As we
find in our
report, the monic FIR noise feedback filters that we choose have
the unfortunate side-effect of boosting truncation noise
in part of the
spectral range.
Hence, truncation error feedback is not effective in reducing bit
rate.
This may help to explain why we have not seen this idea
appearing in the
literature.