Kernel learning for blind image recovery from motion blur

Restoring image from motion deblur faces great challenges in the estimation of the motion blur kernel that is the key to recover the latent sharp image. In this paper, we present a method to iteratively estimate the structural image and account for the textural component. A scale-aware smoothing operation is developed to remove fine-scale edges with resampling. Our method leverages L-0-norm regularization to enforce the sparsity of the motion blur kernel in both intensity and derivative domains. Experiments are conducted to evaluate the performance of our proposed method using two widely accepted public datasets. We found that our proposed method is insensitive to most hyper-parameters. Both qualitative evaluation and quantitative evaluation confirms that our method effectively restores the sharp image without introducing artifacts. The minimum improvements in terms of average PSNR for both datasets are more than 3.13% for all cases and the improvements in terms of average error rate are 15%. By visually comparing the estimated motion blur kernels, it is clear that the estimated kernel by our method is the closest to the actual kernel used to generate the synthesized blurry images.