Figure 4

Statistical distribution of the DCT coefficients of the residual noise plotted in Figure 3 . For comparison, the Laplacian pdf with parameters estimated by the maximum likelihood estimation (ML est.) are also shown in (a). Note that for a meaningful comparison, (b) shows the results after normalization by the estimated scale parameter \(\widehat {b}_{k}\). It is noted that three denoising filter are respectively designed based on wavelet, block-matching and 3D (BM3D) and dictionary learning (K-SVD) algorithms.