Typical blur from camera shake often deviates from the standard uniformconvolutional script, in part because of problematic rotations which creategreater blurring away from some unknown center point. Consequently, successfulblind deconvolution requires the estimation of a spatially-varying ornon-uniform blur operator. Using ideas from Bayesian inference and convexanalysis, this paper derives a non-uniform blind deblurring algorithm withseveral desirable, yet previously-unexplored attributes. The underlyingobjective function includes a spatially adaptive penalty which couples thelatent sharp image, non-uniform blur operator, and noise level together. Thiscoupling allows the penalty to automatically adjust its shape based on theestimated degree of local blur and image structure such that regions with largeblur or few prominent edges are discounted. Remaining regions with modest blurand revealing edges therefore dominate the overall estimation process withoutexplicitly incorporating structure-selection heuristics. The algorithm can beimplemented using a majorization-minimization strategy that is virtuallyparameter free. Detailed theoretical analysis and empirical validation on realimages serve to validate the proposed method.
展开▼
