The central nervous system encodes information in sequences of asynchronously generated voltage spikes, but
the precise details of this encoding are not well understood. Thorpe proposed rank-order codes as an explanation
of the observed speed of information processing in the human visual system. The work described in this paper
is inspired by the performance of SpikeNET, a biologically inspired neural architecture using rank-order codes
for information processing, and is based on the retinal model developed by VanRullen and Thorpe. This model
mimics retinal information processing by passing an input image through a bank of Difference of Gaussian (DoG)
filters and then encoding the resulting coefficients in rank-order. To test the effectiveness of this encoding in
capturing the information content of an image, the rank-order representation is decoded to reconstruct an image
that can be compared with the original. The reconstruction uses a look-up table to infer the filter coefficients
from their rank in the encoded image. Since the DoG filters are approximately orthogonal functions, they are
treated as their own inverses in the reconstruction process.
We obtained a quantitative measure of the perceptually important information retained in the reconstructed
image relative to the original using a slightly modified version of an objective metric proposed by Petrovic. It is
observed that around 75% of the perceptually important information is retained in the reconstruction.
In the present work we reconstruct the input using a pseudo-inverse of the DoG filter-bank with the aim of
improving the reconstruction and thereby extracting more information from the rank-order encoded stimulus.
We observe that there is an increase of 10 - 15% in the information retrieved from a reconstructed stimulus as
a result of inverting the filter-bank.