Deep LoRA-Unfolding Networks for Image Restoration.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep unfolding networks (DUNs) integrate iterative optimization with deep neural networks for image restoration tasks.
• Existing DUNs have limitations in stage-specific noise adaptation and suffer from parameter redundancy, hindering efficiency.

Purpose of the Study:

• To introduce a novel, efficient DUN framework for image restoration.
• To address limitations of parameter redundancy and lack of stage-specific adaptation in existing DUNs.

Main Methods:

• Developed generalized Deep Low-rank Adaptation (LoRA) Unfolding Networks (LoRun) for image restoration.
• LoRun utilizes a shared base denoiser with lightweight, stage-specific LoRA adapters injected into Proximal Mapping Modules (PMMs).
• Dynamically modulates denoising behavior based on noise levels at each unfolding step, decoupling core restoration from adaptation.

Main Results:

• Achieved significant parameter reduction (up to N times for an N-stage DUN) with compressed memory usage.
• Demonstrated on-par or superior performance compared to existing methods across three image restoration tasks.
• Validated the efficiency and effectiveness of the LoRun framework.

Conclusions:

• LoRun offers a more efficient and adaptable approach to deep unfolding networks for image restoration.
• The proposed method effectively addresses parameter redundancy and enhances stage-specific noise adaptation.
• LoRun shows promise for large-scale and resource-constrained image restoration applications.