Blur kernel estimation of noisy-blurred image via dynamic structure prior

An accurate blur kernel is key to blind image deblurring and kernel estimation heavily relies on strong edges in the observed image [1, 2, 3]. Previous methods [4] [5] leveraging image gradient prior with i.i.d statistics can hardly restrict strong edges in a noisy-blurred image, since both noise and strong edges are presented as strong gradients. In this paper, we propose a dynamic structure prior (DSP) to distinguish strong edges from noise by a non local total variation regularization. Our method measures accumulated intensity change of each pixel along all possible directions within its neighborhood. In this case, the structural contents with less change in certain directions, i.e.strong edges, will be accumulated lower than non-structural contents such as noise. By contrast, the previous image gradient priors only measure the changes in horizontal and vertical directions, which is not capable to present the directional characteristic of image edges. Moreover, in our method, the prior changes the weight dynamically with regard to local structures to impose fewer penalties on pixels with large changes. As a result, strong edges are preserved while fine textures are suppressed by large penalties. At last, a cost function containing the dynamic structure prior and L₀ gradient prior that sharpens the preserved edges is proposed to estimate blur kernels from noisy-blurred images. Comprehensive experimental results show that our method outperforms previous methods in both commonly used datasets with various noise levels and real world images, from monochrome image to color image.