"Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration"

The paper introduces a novel domain adaptation method for image restoration.

It leverages diffusion models to bridge the performance gap between synthetic and real-world degraded images. This approach uses a 'diffusion loss' in the noise space to align synthetic and real image distributions.

-----

Paper - https://arxiv.org/abs/2406.18516

Original Problem 😕:

→ Image restoration models trained on synthetic data often fail to generalize to real-world images.

→ This is due to the domain gap between synthetic and real image degradations.

→ Existing domain adaptation methods struggle in low-level vision tasks like image restoration.

-----

Solution in this Paper 💡:

→ This paper proposes "Denoising as Adaptation".

→ It uses a diffusion model to guide a restoration network.

→ The diffusion model is conditioned on both restored synthetic and real-world images.

→ A 'diffusion loss' is derived from the diffusion model's noise prediction error.

→ This loss encourages the restoration network to output high-quality images, aligning both synthetic and real image distributions to a clean target distribution.

→ To prevent shortcut learning, channel shuffling and residual-swapping contrastive learning are introduced in the diffusion model.

→ Channel shuffling randomizes input channels.

→ Residual-swapping contrastive learning ensures the model learns genuine restoration, not just recognizing synthetic data.

-----

Key Insights from this Paper 🤔:

→ Diffusion models can be used for domain adaptation in image restoration by operating in the noise space.

→ The noise prediction error of a diffusion model is sensitive to the quality of conditional inputs.

→ This sensitivity can be leveraged as a 'diffusion loss' to guide domain adaptation.

→ Strategies like channel shuffling and contrastive learning are crucial to prevent shortcut learning in joint training.

-----

Results 📊:

→ Achieves 34.71 PSNR and 0.9202 SSIM on SIDD denoising dataset, outperforming previous domain adaptation methods.

→ Reaches 34.39 PSNR on SPA deraining dataset, surpassing feature and pixel space adaptation techniques.

→ Attains 26.46 PSNR on RealBlur-J deblurring dataset, showing competitive performance.