Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration
Kang Liao, Zongsheng Yue, Zhouxia Wang, Chen Change Loy
2025-01-27
Summary
This paper talks about AdaIR, a new method for fixing various types of image problems all at once. It uses something called frequency mining and modulation to make images clearer and better-looking, no matter what kind of issue they have.
What's the problem?
When we take pictures, lots of things can go wrong. The image might be blurry, noisy, hazy, or have rain in it. Usually, we need different tools to fix each of these problems. Some newer methods try to fix all these issues with one tool, but they don't look at how different problems affect different parts of the image's frequency information. This means they might miss important details that could help fix the image better.
What's the solution?
The researchers created AdaIR, which looks at both the regular image and its frequency information. It does this in three main steps. First, it finds important information in both low and high frequencies of the image. Then, it adjusts this information to make it more useful. Finally, it combines this adjusted information with the original image to make it clearer and better. This method can adapt to different types of image problems because it pays attention to how each problem affects different frequency parts of the image.
Why it matters?
This research matters because it could make it much easier to fix all kinds of image problems with just one tool. Instead of needing separate programs for fixing blurry photos, removing haze, or clearing up dark images, we could have one program that does it all. This could be really helpful for photographers, but also for things like self-driving cars or medical imaging, where clear images are super important. The researchers showed that their method works better than other current methods for lots of different image fixing tasks, which is pretty exciting for anyone who works with or enjoys looking at photos.
Abstract
Although learning-based image restoration methods have made significant progress, they still struggle with limited generalization to real-world scenarios due to the substantial domain gap caused by training on synthetic data. Existing methods address this issue by improving data synthesis pipelines, estimating degradation kernels, employing deep internal learning, and performing domain adaptation and regularization. Previous domain adaptation methods have sought to bridge the domain gap by learning domain-invariant knowledge in either feature or pixel space. However, these techniques often struggle to extend to low-level vision tasks within a stable and compact framework. In this paper, we show that it is possible to perform domain adaptation via the noise space using diffusion models. In particular, by leveraging the unique property of how auxiliary conditional inputs influence the multi-step denoising process, we derive a meaningful diffusion loss that guides the restoration model in progressively aligning both restored synthetic and real-world outputs with a target clean distribution. We refer to this method as denoising as adaptation. To prevent shortcuts during joint training, we present crucial strategies such as channel-shuffling layer and residual-swapping contrastive learning in the diffusion model. They implicitly blur the boundaries between conditioned synthetic and real data and prevent the reliance of the model on easily distinguishable features. Experimental results on three classical image restoration tasks, namely denoising, deblurring, and deraining, demonstrate the effectiveness of the proposed method.