Abstract:All-in-one image restoration tasks are becoming increasingly important, especially for ultra-high-definition (UHD) images. Existing all-in-one UHD image restoration methods usually boost the model's performance by introducing prompt or customized dynamized networks for different degradation types. For the inference stage, it might be friendly, but in the training stage, since the model encounters multiple degraded images of different quality in an epoch, these cluttered learning objectives might be information pollution for the model. To address this problem, we propose a new training paradigm for general image restoration models, which we name \textbf{Review Learning}, which enables image restoration models to be capable enough to handle multiple types of degradation without prior knowledge and prompts. This approach begins with sequential training of an image restoration model on several degraded datasets, combined with a review mechanism that enhances the image restoration model's memory for several previous classes of degraded datasets. In addition, we design a lightweight all-purpose image restoration network that can efficiently reason about degraded images with 4K ($3840 \times 2160$) resolution on a single consumer-grade GPU.

What problem does this paper attempt to address?

### What problems does this paper attempt to solve? This paper aims to solve several key problems in ultra - high - definition (UHD) image inpainting: 1. **A unified model for multi - type degradation processing**: - Existing UHD image inpainting methods usually need to train multiple specialized models for different degradation types (such as denoising, deblurring, deraining, desnowing, low - light enhancement, etc.) respectively. This not only increases the demand for computing resources but also limits the generalization ability of the model. - The paper proposes a general - purpose image inpainting model that can handle multiple types of degradation simultaneously without the need to design specific network architectures or prompt information for each degradation type. 2. **Avoiding information pollution and catastrophic forgetting**: - During the training process, when the model encounters multiple different types of degraded images, it may lead to information pollution, that is, the information of different types of degradation interferes with each other, affecting the learning effect of the model. - At the same time, as the model learns new tasks, it may forget the knowledge it has learned before. This phenomenon is called catastrophic forgetting. - The paper solves these problems by introducing the "Review Learning" method. This method regularly reviews the previously learned difficult samples during the training process to enhance the model's memory and prevent information pollution. 3. **Design of an efficient and lightweight model**: - In order to achieve full - resolution inference of 4K or even 8K resolution images on a single consumer - level GPU, the paper designs a lightweight and efficient general - purpose image inpainting model - SimpleIR. This model can significantly reduce the number of parameters while maintaining high performance, thereby reducing the demand for computing resources. ### The core idea of Review Learning - **Review learning mechanism**: By gradually training the model to process different types of degraded data and reviewing the previously learned difficult samples after each stage, it is ensured that the model will not forget the learned knowledge and can effectively deal with new degradation types. - **Sample selection based on entropy difference**: By analyzing the entropy difference of samples, the most challenging samples are identified for review, avoiding simply relying on loss values to select difficult samples, thus more accurately capturing the samples that are difficult for the model to handle. - **Lightweight network structure**: A lightweight network architecture, SimpleIR, is designed, which can efficiently process high - resolution images on a single consumer - level GPU while maintaining high inpainting performance. ### Experimental results The paper conducts experiments on multiple benchmark datasets to verify the effectiveness of the proposed method. The experimental results show that SimpleIR has achieved significantly better performance than existing methods in multiple tasks such as desnowing, deblurring, deraining, and low - light enhancement, and has fewer parameters and higher computational efficiency. In conclusion, through proposing the Review Learning method and the SimpleIR model, this paper successfully solves the problems of multi - type degradation processing, information pollution, and catastrophic forgetting in UHD image inpainting, providing a new solution for achieving efficient and general - purpose image inpainting.

Review Learning: Advancing All-in-One Ultra-High-Definition Image Restoration Training Method

Model Parameter Learning for Real-Time High-Resolution Image Enhancement.

A Comprehensive Review of Deep Learning-Based Real-World Image Restoration

Multi-modal Degradation Feature Learning for Unified Image Restoration Based on Contrastive Learning

A survey of deep learning approaches to image restoration

A Double Feature Fusion Network with Progressive Learning for Sharper Inpainting

All-in-one Multi-degradation Image Restoration Network via Hierarchical Degradation Representation

OneRestore: A Universal Restoration Framework for Composite Degradation

Restorer: Removing Multi-Degradation with All-Axis Attention and Prompt Guidance

Image Restoration using Feature-guidance

A Method for Remote Sensing Image Restoration Based on the System Degradation Model

Perceive-IR: Learning to Perceive Degradation Better for All-in-One Image Restoration

Chain-of-Restoration: Multi-Task Image Restoration Models are Zero-Shot Step-by-Step Universal Image Restorers

A Survey on All-in-One Image Restoration: Taxonomy, Evaluation and Future Trends

Multistage supervised contrastive learning for hybrid-degraded image restoration

Neural Degradation Representation Learning for All-In-One Image Restoration

In-training Restoration Models Matter: Data Augmentation for Degraded-reference Image Quality Assessment

Adaptive Blind All-in-One Image Restoration

Unified-Width Adaptive Dynamic Network for All-In-One Image Restoration

A Progressive Image Restoration Network for High-order Degradation Imaging in Remote Sensing

Reference-based Multi-stage Progressive Restoration for Multi-degraded Images