Abstract:Deep neural networks (DNNs) have achieved significant success in image restoration tasks by directly learning a powerful non-linear mapping from corrupted images to their latent clean ones. However, there still exist two major limitations for these deep learning (DL)-based methods. Firstly, the noises contained in real corrupted images are very complex, usually neglected and largely under-estimated in most current methods. Secondly, existing DL methods are mostly trained on one pre-assumed degradation process for all of the training image pairs, such as the widely used bicubic downsampling assumption in the image super-resolution task, inevitably leading to poor generalization performance when the true degradation does not match with such assumed one. To address these issues, we propose a unified generative model for the image restoration, which elaborately configures the degradation process from the latent clean image to the observed corrupted one. Specifically, different from most of current methods, the pixel-wisely non-i.i.d. Gaussian distribution, being with more flexibility, is adopted in our method to fit the complex real noises. Furthermore, the method is built on the general image degradation process, making it capable of adapting diverse degradations under one single model. Besides, we design a variational inference algorithm to learn all parameters involved in the proposed model with explicit form of objective loss. Specifically, beyond traditional variational methodology, two DNNs are employed to parameterize the posteriori distributions, one to infer the distribution of the latent clean image, and another to infer the distribution of the image noise. Extensive experiments demonstrate the superiority of the proposed method on three classical image restoration tasks, including image denoising, image super-resolution and JPEG image deblocking.

Unrolling Plug-and-Play Gradient Graph Laplacian Regularizer for Image Restoration

Deep Generalized Unfolding Networks for Image Restoration

Interpretable Lightweight Transformer via Unrolling of Learned Graph Smoothness Priors

Image Restoration Based on End-to-End Unrolled Network

Constructing an Interpretable Deep Denoiser by Unrolling Graph Laplacian Regularizer

Unrolling of Deep Graph Total Variation for Image Denoising

Graph Unrolling Networks: Interpretable Neural Networks for Graph Signal Denoising

Deep Graph Laplacian Regularization for Robust Denoising of Real Images

Deep, convergent, unrolled half-quadratic splitting for image deconvolution

Variational Image Restoration Network

Variational Image Restoration Network.

Iterative deep neural networks based on proximal gradient descent for image restoration

Graph Signal Restoration Using Nested Deep Algorithm Unrolling

Deep unfolding of a proximal interior point method for image restoration

Graph Signal Denoising Via Unrolling Networks.

Retinex-based Image Denoising / Contrast Enhancement using Gradient Graph Laplacian Regularizer

CurvPnP: Plug-and-play Blind Image Restoration with Deep Curvature Denoiser

Learning Deep Gradient Descent Optimization for Image Deconvolution

Manifold Graph Signal Restoration Using Gradient Graph Laplacian Regularizer