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.

Two Matrix-Type Projection Neural Networks for Matrix-Valued Optimization with Application to Image Restoration

Analysis and Application of Matrix-Form Neural Networks for Fast Matrix-Variable Convex Optimization

Non-Local Image Inpainting Using Low-Rank Matrix Completion: Non-Local Image Inpainting Using Low-Rank Matrix Completion

A discrete-time neural network for optimization problems with hybrid constraints.

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

A New Delayed Projection Neural Network for Solving Linear Variational Inequalities and Quadratic Optimization Problems

Discrete-Time Projection Neural Network Methods for Computing the Solution of Variational Inequalities

A Double Feature Fusion Network with Progressive Learning for Sharper Inpainting

A modified projection neural network for linear variational inequalities and quadratic optimization problems

Design of General Projection Neural Networks for Solving Monotone Linear Variational Inequalities and Linear and Quadratic Optimization Problems

A collaborative neurodynamic approach with two-timescale projection neural networks designed via majorization-minimization for global optimization and distributed global optimization

Discrete-Time Neural Network for Fast Solving Large Linear $L_{1}$ Estimation Problems and Its Application to Image Restoration

Deep Block Proximal Linearised Minimisation Algorithm for Non-convex Inverse Problems

Solving Pseudomonotone Variational Inequalities and Pseudoconvex Optimization Problems Using the Projection Neural Network.

A generalized global convergence theory of projection-type neural networks for optimization

Study Of Neural Networks In Image Restoration

A Novel Projection Neural Network for Solving a Class of Monotone Variational Inequalities.

On exponential convergence conditions of an extended projection neural network

Optimization on Matrix Manifold Based on Gradient Information and Its Applications in Network Control

Variational Image Restoration Network.

Non-Local Auto-Encoder With Collaborative Stabilization for Image Restoration.