An Integrated Framework for Image Classification
X. Luo and G. Mirchandani
ICASSP2000
(Paper (pdf) / Paper (ps))
This paper presents a novel method for classifying an image into
one of predefined classes in a data bank by introducing the
application of mutual information in the Fourier amplitude
domain. Template and test images are made translation and rotation
invariant through the Fourier-Mellin transform. While mutual information
could be employed here, we choose instead to apply it to a lower
dimension phase spectrum generated by the complex multiresolution
wreath product transform of the Fourier amplitude spectrum.
The phase information of this transform adequately preserves edges
even at low resolutions while permitting, at the same time, a
reduction in the computational burden. Brodatz textures and ORL faces
are used to demonstrate the capability of this algorithm.