Blind deconvolution using a nondimensional Gaussianity measure

Blind image deconvolution (BID) is a severely ill-posed problem which requires prior information on the latent image to estimate the blur kernel. In this paper, a new observation that blurring always pushes the gradient of a local image region toward its mean value is introduced. And we formulate a novel function to measure the distance between the local gradient and its mean value. A novel regularizer associated with local gradient means is proposed. As it requires to segment the whole image into small regions, we propose an approximate method without any segmentation. Thanks to its simplicity the algorithm is fast and robust. Numerous experimental results on synthetic and real data demonstrate that our method is capable of removing various uniform blurs such as motion blur, atmospheric blur and out-of-focus blur.