Single Image Blind Deblurring With Local Maximum Gradient Prior

Single image blind deblurring,as one of the most intriguing problem in the area of image restoration,has important research significance and practical application value.Blind deblurring often involves the process of kernel estimation and non-blind restoration.Existing blur kernel estimation methods have their limitations as they are not able to handle blurry images under complicate situations.Moreover,recent nonblind deconvolution algorithms are often sensitive to outliers,which will ultimately result in terrible final restoring results.Our contributions can be summarized as followings,(1)We propose a new local maximum gradient prior,which can help improve the quality of estimated blur kernel.Specifically,we find that the maximum gradient value of a single image patch tends to diminish after the blurring process.We investigate this appearance with a sound theoretical explanation.By incorporating the proposed prior into a conventional deblurring framework,we manage to acquire a high-quality blur kernel.During the optimization process,we introduce a linear operator to compute the maximum gradient value of a local patch.In a half quadratic splitting optimization framework,our minimization scheme convergences quickly in practice.(2)After obtaining the blur kernel,we propose to restore the final sharp image with a robust non-blind deblurring method which is based on maximum Gini impurity.In practice,blurry images often contain lots of outliers such as impulsive noise and saturated pixels,which do not comply with the linear convolution rule of blurring.Thus,we are unable to recover sharp images with outliers given the exactly blur kernel.Our method can recover blurry images by classifying and eliminating outliers during recovering.We analyse the proposed method on different provided benchmark datasets against state-of-the-art methods.Our algorithm performs favorably not only in synthetic images but natural images under different conditions,such as text,face and lowillumination blurry images.Further empirically analysis the reason of the superiority of the proposed prior over other image priors.