Adaptive Optimization with Nested Prior Navigation for Blind Image Deblurring

Abstract

Image deblurring is a typical inverse problem and has many applications in vision and multimedia areas. It involves the estimation of blur kernel and sharp image given only a blurry observation. Due to the fundamentally ill-posed nature, most existing works design different priors within the maximum a posteriori (MAP) framework to regularize their solution space. However, due to the unknown image distribution, complex kernel structure and non-uniform noises, it is indeed challenging to explicitly design a fixed prior for blurry images in realworld scenarios. Different from these conventional strategies to integrate sophisticated priors into optimization model, this work only formulates the necessary constraints on latent image and blur kernel as a lightweight MAP model. Then we develop an inexact projected gradient scheme to incorporate flexible sparse structure control for MAP inference. We demonstrate that this adaptive scheme can successfully avoid degenerate solutions and is universally suitable for different blurry scenarios, such as low-illumination, face and text. Extensive experiments on both synthetic data and real-world images demonstrate the effectiveness and robustness of the proposed method against other state-of-the-art methods.