It has been proven that structure of image play an important role in kernel estimation.In recent years

many successful algorithms propose to generate intermediate image by extracting structure from latent image

and then use it for blur kernel estimation.However

these methods ignore to extract the correspondence from input blurry image.This will cause unbalanced data item of objective function.In this paper we first exploit a mask determined by convolution of intermediate image with kernel to generate the correspondence

and then take it into data item instead of blurry image to overcome the problem.Moreover

we have found that kernel shows the properties of sparse both in intensity domain and derivatives domain.Accordingly

we apply L0-norm regularization to constrain both intensity domain and derivatives domain of kernel.Compared with the state-of-the-art algorithms

experiments across datasets showed that our algorithm achieved better performance.