Figure 13.

Blind deconvolution of five different realistic natural images. From left to right: (a) input blurred images, deblurred versions generated by (b) Xu and Jia [11], (c) Krishnan et al. [9], (d) Pan et al. [59] and (e) our proposed method, respectively. The sizes of the estimated blur kernels from top to bottom are $35\times 35$, $55\times 55$, $41\times 41$, $35\times 35$ and $55\times 55$, respectively (the images are best viewed in full-screen mode).