Variational Diffusion Method for Blind Image Deblurring

Since the images generated by the diffusion model have outstanding visual effects, it has been widely used in image deblurring, primarily when the blur kernel is known. However, its application in blind image deblurring is still less and needs further exploration. Image deblurring can be formulated as inferring the posterior distribution of the clear image with a given blurred image. Although there are several methods for estimating the posterior distribution, they are generally computationally inefficient and time-consuming. In this paper, we propose a variational method to approximate the true posterior distribution while combining the Expectation-Maximization algorithm to introduce the update of the blur kernel into the inverse diffusion process. The clear image is obtained by alternating iterations of diffusion model and kernel estimation. Experimental results validate the effectiveness and efficiency of the proposed method when compared to the state-of-the-art approaches based on diffusion models.