Modeling Spatial and Temporal Textures
Bridging the gap between the Wold random process decomposition theory
and practical texture modeling, this research establishes Wold-based
texture modeling as an important method for a wide range of
applications that benefit from efficient and effective
characterization of textural information.
A robust and efficient algorithm is developed for spectral 2-D Wold
decomposition of homogeneous or near homogeneous random fields.
A psychophysical study is conducted to show that the Wold component
energy of a texture pattern is a good computational measure for the
most salient human texture perception dimension of repetitiveness vs.
A shift, rotation, and scale invariant Wold-based texture model is
presented. This model provides efficient and perceptually sensible
features that are robust to many natural texture inhomogeneities.
For model perspective invariance, a linear system characterization and
a decomposition of image perspective transformations are proposed to
form a basis for future algorithms to infer image perspective
parameters from a single sample of harmonic texture data.
Based on the Wold texture model, an algorithm is developed for
textured image database retrieval. Compared to the state-of-the-art
texture models, the new model appears to offer perceptually more
satisfying retrieval results while matching or surpassing the best
recognition performance of the others.
A K-means-based image segmentation method is presented to demonstrate
the use of Wold-based modeling in characterizing textured regions in
natural scene images.
Applying the principle of Wold decomposition to temporal texture
modeling, a robust and efficient algorithm is developed for detecting
and segmenting periodic motion. The use of periodicity templates is
proposed for characterizing periodicity in space and time.
Thesis Supervisor: Rosalind W. Picard
Title: NEC Development Professor of Computers and Communications
Compressed Postscript .
Full list of tech reports