Abstract
As a classical image processing problem, image denoising arms to estimate the original image from the noisy version according to the prior conditions. Recent studies have shown that deep neural networks can also be seen as an image prior, referring to the Deep Image Prior (DIP) technique. The DIP method is able to recover a natural image only using random inputs and the corrupted observation through performing correction via a convolutional network. In this paper, we propose to improve the performance of the DIP method. Firstly, the original optimization objective function is modified by adding two nonlocal regularizers to promote sparsity of the solution. Secondly, we solve the optimization problem with the Alternating Direction Method of Multipliers (ADMM) framework, leading to a plug-and-play ADMM method for deep learning-based image denoising. Experiments validate the effectiveness of leveraging a combination of DIP and nonlocal regularizers, and demonstrate the superior performance of our method both quantitatively and visually as compared with the original DIP method.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call