Unified Image Restoration and Enhancement: Degradation Calibrated Cycle Reconstruction Diffusion Model

TL;DR

CycleRDM addresses the challenge of unifying image restoration and enhancement under real-world, diverse degradations by introducing a degradation-calibrated, multi-stage diffusion framework. It learns progressive mappings from degraded to rough normal to normal domains, then transfers calibration to the wavelet low-frequency domain and refines high-frequency details with a feature gain module, all guided by multimodal prompts and frequency-domain losses. The approach achieves strong performance across nine tasks with limited training data, demonstrated through extensive quantitative and qualitative evaluations and comprehensive ablations. This method offers a practical, generalizable pathway for robust, high-quality image recovery and enhancement in real-world settings, with publicly available code to enable replication and extension.

Abstract

Image restoration and enhancement are pivotal for numerous computer vision applications, yet unifying these tasks efficiently remains a significant challenge. Inspired by the iterative refinement capabilities of diffusion models, we propose CycleRDM, a novel framework designed to unify restoration and enhancement tasks while achieving high-quality mapping. Specifically, CycleRDM first learns the mapping relationships among the degraded domain, the rough normal domain, and the normal domain through a two-stage diffusion inference process. Subsequently, we transfer the final calibration process to the wavelet low-frequency domain using discrete wavelet transform, performing fine-grained calibration from a frequency domain perspective by leveraging task-specific frequency spaces. To improve restoration quality, we design a feature gain module for the decomposed wavelet high-frequency domain to eliminate redundant features. Additionally, we employ multimodal textual prompts and Fourier transform to drive stable denoising and reduce randomness during the inference process. After extensive validation, CycleRDM can be effectively generalized to a wide range of image restoration and enhancement tasks while requiring only a small number of training samples to be significantly superior on various benchmarks of reconstruction quality and perceptual quality. The source code will be available at https://github.com/hejh8/CycleRDM.

Figures (8)

• Figure 1: CycleRDM is capable of generating high-fidelity restoration in a variety of tasks. CycleRDM gives faithful results on a wide range of (a) linear image restoration tasks. Meanwhile, CycleRDM also realizes (b) blind, non-linear image enhancement tasks with high quality.
• Figure 2: The proposed overall framework for CycleRDM. We use image deblurring as a demonstration. Firstly, in Stage 1 we use the degraded image $LQ$ as a condition guidance to learn the mapping relation between the degraded domain to the rough normal domain, and later to guide the learning of the rough normal domain to the normal domain in Stage 2. In Stage 3, we perform a discrete wavelet transform ($DWT$) on the output $\hat{x}^{2}_{0}$ of stage 2. At the same time, a fine calibration is performed in the wavelet low-frequency domain $L$ using the degradation prior learned earlier. For each Stage output, we also utilise multimodal text for appearance guidance. And the high-frequency $H$ is enhanced by the feature gain module ($FGM$), and finally recovered to a high-quality image $HQ$ by the inverse discrete wavelet transform ($IDWT$).
• Figure 3: Comparison of our method with other methods on 4 different degradation-specific tasks. Where the first to fourth rows are dehazing, denoising, deblurring, and deraining respectively. Best viewed by zooming in.
• Figure 4: Comparison of our method with other state-of-the-art methods in image restoration tasks with different degradation types. Best viewed by zooming in.
• Figure 5: Comparison of our method with competing methods on low light image enhancement task. Best viewed by zooming in.