Multiresolution Texture Segmentation with the DFT-based Complex
Lapped Transform Phase
X. Luo and G. Mirchandani
2001 IEEE-EURASIP Workshop on Nonlinear Signal and Image Processing
(Paper (pdf) / Paper (ps))
In this paper we generalize the lapped orthogonal transforms to include
complex filter coefficients. We derive the perfect reconstruction conditions
for this class of complex lapped transforms. A special discrete Fourier
transform-based complex lapped transform (DCLT) is constructed which
tends to provide a good mixture of locality and smoothness. In applying
the DCLT to an image, a quadtree scan is employed following the
wreath product transform style. We apply the DCLT to texture segmentation.
A multiresolution algorithm is proposed, which uses the transform phase
to represent the higher scale image data, and uses the Markov random field
model to represent the segmentation labels.
Performance on texture segmentation with the DCLT is compared
to that with the usual lowpass coefficients representation.
When the texture to be discriminated has major local edge
directions, the segmentation results appear to conform to our
our visual systems.