These examples illustrate blind deconvolution on synthetically blurred noisy images created from a combination of eight unblurred test images and eight blur kernels. There are three versions of the synthetic images for each combination, differing in how much Poisson noise is added: None, S = 100 (less noisy), and S = 10 (more noisy). The blur kernel is estimated from each input image as a preliminary step, using the method of Levin et al (2011), Efficient Marginal Likelihood Optimization in Blind Deconvolution. Results for all eight kernels are shown together; the true (left) and estimated (right) kernels are shown above each image.
Results are shown for an IID Gaussian prior (α = 2, λ = 0.01), an IID sparse prior (α = 0.8, λ = 0.001), and an AR sparse prior (α = 0.8, λ = 0.001) which additionally assumes that the image gradients are correlated (according to a simple independent 2-D auto-regressive model with correlation parameters 0.3 and 0.6). For comparison, results using the Richardson-Lucy algorithm (after 25 iterations) and the non-blind deconvolution method used by Levin et al are also shown; the latter is essentially a version of the IID sparse prior that regularizes using a third second-order gradient in addition to the horizontal and vertical gradients.
Non-blind deconvolution on the same set of images (using the true kernel) is shown here.