An elliptically contoured exponential mixture model for wavelet based image denoising (Q2371331): Difference between revisions
From MaRDI portal
Changed an Item |
Changed an Item |
||
Property / describes a project that uses | |||
Property / describes a project that uses: DT-CWT / rank | |||
Normal rank |
Revision as of 02:04, 28 February 2024
scientific article
Language | Label | Description | Also known as |
---|---|---|---|
English | An elliptically contoured exponential mixture model for wavelet based image denoising |
scientific article |
Statements
An elliptically contoured exponential mixture model for wavelet based image denoising (English)
0 references
4 July 2007
0 references
In earlier work of Sendur and Selesnick, univariate and bivariate wavelet shrinkage models for denoising were proposed to exploit interscale dependency. The present paper extends those methods to spherically and elliptically contoured exponential models (SCEM and ECEM) used as priors for expectation maximization algorithms for image denoising. The two component SCEM is defined as \[ f_X^M(x) = rf_X(x| \sigma_{x1})+(1-r)f_X(x| \sigma_{x2}) \] where \(f(x| \sigma_{xi})\) are SCE distributions with different variances and \(0\leq r\leq 1\) is a ratio parameter. The idea is to capture the peak of a distribution of wavelet coefficients with the component having small variance and to capture the tail with that having large variance. For this mixture model there is no closed form solution for the maximum likelihood estimator so instead the authors propose the following algorithm for estimating \(r,\sigma_{x1}\) and \(\sigma_{x2}\) based on the hypothesis that the data fits the SCE model: (i) Initialize \(r,\sigma_{x1},\sigma_{x2}\) and repeat (ii) and (iii) until convergence, where (ii) involves calculating for each sample \(x_m\in \mathbb{R}^d\), \(R_m=rf_X(x_m| \sigma_{x1})/f_X^{\text{est}}(x_m)\) where \(f_X^{\text{est}}\) is the current estimate of \(f_X^M\), then updating \(r\) by the running average of the \(R_m\), and (iii) updating \(\sigma_{x1}\) as the maximizer of \(\sum \log [f_X(x_m| \sigma_{x1})]R_m\) and \(\sigma_{x2}\) as the maximizer of \(\sum \log [f_X(x_m| \sigma_{x2})](1-R_m)\). Similar techniques are used for the more general ECE models that allow for correlations across wavelet coefficient scales. Results comparing mixtures of Gaussian and Laplacian distributions, and non mixture models are presented and compared to other wavelet coefficient denoising methods.
0 references
expectation maximization
0 references
wavelet denoising
0 references