UIR-LoRA: Achieving Universal Image Restoration through Multiple Low-Rank Adaptation

Abstract: Existing unified methods typically treat multi-degradation image restoration as a multi-task learning problem. Despite performing effectively compared to single degradation restoration methods, they overlook the utilization of commonalities and specificities within multi-task restoration, thereby impeding the model's performance. Inspired by the success of deep generative models and fine-tuning techniques, we proposed a universal image restoration framework based on multiple low-rank adapters (LoRA) from multi-domain transfer learning. Our framework leverages the pre-trained generative model as the shared component for multi-degradation restoration and transfers it to specific degradation image restoration tasks using low-rank adaptation. Additionally, we introduce a LoRA composing strategy based on the degradation similarity, which adaptively combines trained LoRAs and enables our model to be applicable for mixed degradation restoration. Extensive experiments on multiple and mixed degradations demonstrate that the proposed universal image restoration method not only achieves higher fidelity and perceptual image quality but also has better generalization ability than other unified image restoration models. Our code is available at https://github.com/Justones/UIR-LoRA.

• The paper introduces a universal image restoration framework that employs multiple low-rank adapters to tackle various degradation tasks.
• It utilizes a unique LoRA composing strategy based on degradation similarity to adaptively merge task-specific features.
• Extensive experiments show improved fidelity and generalization, offering scalable and efficient solutions for diverse image restoration challenges.

The paper "UIR-LoRA: Achieving Universal Image Restoration through Multiple Low-Rank Adaptation" introduces a novel framework for tackling the problem of image restoration across multiple types of degradation. Unlike traditional methods that address multi-degradation image restoration as a multi-task learning problem, this work emphasizes the integration of both commonalities and specificities across different tasks to enhance performance.

Key Contributions

1. Universal Framework for Image Restoration: The authors propose a universal image restoration framework that employs multiple low-rank adapters (LoRA) drawn from multi-domain transfer learning strategies. This approach seeks to leverage a shared pre-trained generative model to manage various degradation tasks effectively.
2. Low-Rank Adaptation (LoRA): The framework utilizes low-rank adaptation mechanisms to transfer learned features from pre-trained models to specific degradation restoration tasks. This adaptation allows exploiting latent common structures while being tailored to particular degradation types.
3. LoRA Composing Strategy: A unique aspect of this framework is a composing strategy based on degradation similarity. This strategy adaptively combines pre-trained LoRAs, enabling the model to handle mixed degradation scenarios, thus enhancing its versatility and applicability.
4. Performance and Generalization: Through extensive experiments, the method is shown to achieve superior fidelity and perceptual image quality compared to other unified restoration models. The approach also demonstrates improved generalization abilities across varied restoration tasks.

Implications and Future Work

• Efficiency and Scalability: By focusing on adaptation via low-rank structures, the method implies potential improvements in computational efficiency and scalability, especially as the complexity of image degradations increases.
• Application Potential: The successful implementation of a universally adaptable restoration framework could lead to advancements in fields requiring robust image processing, such as medical imaging, remote sensing, and digital forensics.

The framework’s effectiveness in handling both multiple and mixed degradations positions it as a significant contribution to the field of image restoration, pushing forward the boundaries of how multi-degradation scenarios are addressed. The availability of the source code on GitHub suggests a commitment to open science, enabling further research and practical applications.