A new complexity prior for multiresolution image denoising

Application of the minimum description length (MDL) principle to multiresolution image denoising has been somewhat unsuccessful to date. This disappointing performance is due to the crudeness of the underlying prior image models, which lead to overly sparse solutions. We propose a new family of complexity priors based on Rissanen's (1984, 1992) universal prior for integers, which produces estimates with better sparsity properties. This method vastly outperforms previous MDL schemes and is competitive with Bayesian estimators using generalized Gaussian priors on wavelet coefficients.