Title: An Analysis of Mammalian Visual System, Statistical
Structures of Natural Scen es and Scale-invariant, Self-similar
Presented by: Gokce Dane
Abstract: Natural images contain some characteristic statistical
regularity that set them apart from purely random images.
Understanding these regularities enable natural images to
be coded more efficiently. Suggesting that a good objective
for an efficie nt coding of natural scenes is to maximize
the sparseness of the representation, it c an be shown that
wavelet-like transforms are very effective in coding natural
images, since they can produce sparse, informative representation
of natural scenes.
In this study, we look at the question that why our visual
system uses a wavelet-like strategy to represent the visual
environment. We show that natural scenes are approx imately
scale invariant with regards to their power and their phase
spectra. Principa lly because of their phase spectra, wavelet-like
transforms are capable of producing sparse, informative representation
of these images. It is suggested that self-similar codes like
wavelet are effective for many natural phenomena since these
phenomena sh ow similar structures to those in natural scenes.