Collaborators:
Haim Permuter (Ben-Gurion University, Israel), Joseph
Francos (Ben-Gurion University, Israel).
Key words:
texture, colour, classification, Gaussian mixture models,
expectation maximization (EM), wavelets.
Resume:
This work uses Gaussian mixture models of image features
to characterize texture and colour properties, and hence to classify image
blocks. The basic quantity in which we are interested is the likelihood of
an image block given its class. Texture models and colour models are
developed separately, and then combined using a decision-theoretic
criterion.
The texture GMMs were tested using a variety of
"structure" features: the distribution of energy in the subbands
of various wavelet decompositions of the block; the energy in subbands of
the DCT of the block; and the parameters of AR models of the block of
different orders. Wavelets were found to perform best, although in
practice little difference was observed on using different mother wavelets
or the DCT with dyadic subbands. The colour GMMs used the RGB mean and
covariance on the block.
The parameters of these models were learned using the EM
algorithm, whose update step can be given in closed form for GMMs.
The ability of the models to retrieve textures from the
VisTex database has been compared to other classic methods on the same
database, and a net improvement observed. Current work is focused on the
retrieval of images from a database of aerial images kindly provided by
the IGN (French National Geographic Institute).
Results:
|
