In this work we present experimental results on the use of the Bispect
rum for translational invariant image averaging. The main issue is to
systematically explore the applicability of Bispectral averaging to pr
actical cases where the starting images are rather noisy. The main con
clusion is that this type of averaging process presents a number of in
trinsic instabilities that makes the whole approach difficult to use.
We start by proving that the sum of two Bispectra is not necessarily a
Bispectrum. We then present a number of practical studies on how many
Bispectra from noisy realizations of a given image are needed for a m
eaningful image recovery as a function of the image signal-to-noise ra
tio. The question of averaging a non-homogeneous data set is also addr
essed. Some practical instabilities associated with recursive recovery
methods are discussed, as well as some practical clues to overcome th
em.