Digital Signal Processing Reference
In-Depth Information
Original
Masked 20% missing pixels Inpainted MCA (Alg. 31) PSNR=35.3 dB
Masked 50% missing pixels Inpainted MCA (Alg. 31) PSNR=31.7 dB
Masked 80% missing pixels Inpainted MCA (Alg. 31) PSNR=26.5 dB
Figure 8.13. Inpainting the “Barbara” image. See Table 8.2 for
the experimental details.
Lines + Gaussians
We repeated the same comparison with a synthetic image, which is a composite of
three Gaussians and three lines; see Fig. 8.15 (top left). On the basis of this morpho-
logical content, the UWT and the curvelet transforms were chosen as candidate dic-
tionaries. The parameters chosen for each algorithm are summarized in Table 8.2.
The masked and filled-in images are portrayed in Fig. 8.15. Both inpainting algo-
rithms performed well, although the result of Algorithm 33 is somewhat smoother.
