We examine the similarity and difference between sparsity and statistical
independence in image representations in a very concrete setting: use the best
basis algorithm to select the sparsest basis
and the least statistically-dependent basis from basis dictionaries for
a given dataset. In order to understand their relationship, we use
synthetic stochastic processes (e.g., spike, ramp, and generalized Gaussian
processes) as well as the image patches of natural scenes.
Our experiments and analysis so far suggest the following:
1) Both sparsity and statistical independence criteria selected similar
bases for most of our examples with minor differences;
2) Sparsity is more computationally and conceptually feasible as a basis
selection criterion than the statistical independence,
particularly for data compression;
3) The sparsity criterion can and should be adapted to individual realization
rather than for the whole collection of the realizations to achieve the maximum
performance;
4) The importance of orientation selectivity of the local Fourier and
brushlet dictionaries was not clearly demonstrated due to the boundary effect
caused by the folding and local periodization.
These observations seem to encourage the pursuit of sparse representations
rather than that of statistically independent representations.
Get the full paper: PDF file.
Get the official version via doi:10.1117/12.408635.
Please email
me if you have any comments or questions!
Go
back to Naoki's Publication Page