Uncertainty Aware Kernel Estimation and Edge Preserving Attention for Blind Image Deblurring

Blind deconvolution is a challenging problem because of its ill-posed nature. Most existing blind deconvolution techniques are based on classical methods and utilize a maximum a posterior (MAP) framework to estimate clean images and blur kernels. Very recently, a method that utilizes the Deep Image Prior (DIP) principle has been proposed. This method uses two generative networks to model the deep priors of clean image and blur kernel. But this method fails for complex kernels, and estimates erroneous kernels, hence leading to ringing artifacts in the reconstructed image. To address this issue and estimate better kernels, we introduce a Bayesian uncertainty guided kernel estimation technique. Also, to improve the quality of the reconstructed images, we present a new type of edge-preserving attention. We perform evaluations on several benchmark datasets to show the performance improvement obtained by our network.