Assessing MDL in wavelet denoising and compression
Mark H Hansen, Bin Yu · 2003
In the context of wavelet denoising and compression, we study two mixture forms of MDL model selection criteria. We find that gMDL outperforms other techniques for 1-dim curve estimation and MDL does the same for natural images. While achieving similar (if not better) MSEs, they keep much fewer coefficients. Hence they are excellent simultaneous denoisers and compressors.