Deconvolution on synthetically blurred images

Reset to Original vs Set as Left Image

Ground truth Richardson-Lucy IID Gaussian AR Gaussian IID Sparse AR Sparse λ=0.01 λ=0.001

These results illustrate the process of non-blind deconvolution on synthetically blurred images created from a combination of eight unblurred test images and eight blur kernels. As the individual images are small, results for all eight images are shown together for each kernel, along with the kernel itself. There are two versions of the synthetic images for each combination, differing in how much Poisson noise is added: S = 100 is less noisy and S = 10 is more noisy.

λ is the regularization parameter; higher values give more importance to the prior. The AR priors assume that the image gradients are correlated (according to a simple independent 2-D auto-regressive model with correlation parameters 0.3 and 0.6). The IID prior assumes they are independent. Results using the classical Richardson-Lucy algorithm (after 25 iterations) is also shown for comparion.

Results are shown using both a direct (sparse matrix) solver and an iterative (conjugate gradient) solver run using 100 iterations for each invocation. Results are similar in most cases, except perhaps a few minor artifacts with some kernels (e.g., the fourth).

The first four test images and all eight test kernels are from Levin et al (2009), Understanding and evaluating blind deconvolution algorithms. The other four test images are from the personal collection of the authors.