Multi-scale Adaptive Image Representations

Doru C. Balcan,
Carnegie Mellon UniversityAbstract:
Multi-scale representations, such as those based on wavelets, have been successful in efficiently
describing intrinsic structure in images, and consequently have lead to the emergence of excellent
image coders. Nevertheless, using any fixed representation limits the best achievable performance
when encoding images of a given class, because all statistical information about that class is simply
ignored. An adaptive representation would then be more appropriate in this setting. One such
example is independent component analysis (ICA), a statistical method that computes a linear
representation whose coefficients have minimum entropy.
In this talk, I will introduce a hybrid image representation method called Multi-scale ICA, which
derives an adaptive basis for each of the wavelet decomposition sub-bands. A direct comparison of
the rate-distortion curves obtained by coefficient scalar quantization to various levels of precision
shows the improvement in terms of efficiency over the wavelet representation. One other merit of this
approach is its potential use to derive adaptive representations for large-size images, where existing
methods fail because of computational and sample complexity limitations. We can therefore interpret
the proposed method both as a nonparametric adaptive extension of wavelet representations, and as
a multi-scale generalization of ICA. This is joint work with Michael Lewicki.Speaker Bio:
Doru C. Balcan received the B.S. degree in computer science, in 2000, and the M.S. degree in
applied computer science, in 2002, from the Faculty of Mathematics, University of Bucharest,
Romania. He is currently pursuing Ph.D. studies in computer science at Carnegie Mellon University,
Pittsburgh, PA. His research is focused on developing algorithms for efficient and robust signal
processing and coding. More exactly, he is interested in techniques that exploit the mathematical
structure of problems commonly occurring in image and audio encoding, to produce representations
that are compact, resilient to noise, and fast to compute.Hosted by: Prof: C-C Jay Kuo

Location: Hughes Aircraft Electrical Engineering Center (EEB) - 248

Audiences: Everyone Is Invited

Contact: Talyia Veal