Figure 5
Comparison of reconstructions of the 3-D BrainWeb phantom corrupted with streak and Poisson noise as in (3), corresponding to the sinograms shown in Fig. 4. Top: ground truth volume, and reconstructions of ln(A) (3) and Z (4) obtained through filtered back-projection. Bottom: comparison of TV reconstruction of estimates with various regularization strengths r, where the percentage implies a multiplier to the regularization optimized to maximize SNR, i.e. 100% means `SNR-optimal' regularization. Top-to-bottom: proposed full estimate ln(A), proposed streak-free estimate , and streak-free estimate of (Vo et al., 2018), each with 100%, 50%, and 150% relative regularization strengths. Proposed estimates are computed embedding the improved BM4D. Notably, the full filtering offers improved reconstruction quality, and is also less sensitive to variations in the regularization parameters. |