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

Authors: Kang Liao, Zongsheng Yue, Zhouxia Wang, Chen Change Loy

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on three classical image restoration tasks, namely denoising, deblurring, and deraining, demonstrate the effectiveness of the proposed method. To verify the effectiveness of our method, we conducted extensive experiments on three classical image restoration tasks (denoising, deblurring, and deraining), showing promising restoration performance and scalability to different networks.
Researcher Affiliation Academia S-Lab, Nanyang Technological University EMAIL
Pseudocode Yes Algorithm 1 Training Process of The Proposed Framework Algorithm 2 One-Pass Inference of The Proposed Framework
Open Source Code Yes https://kangliao929.github.io/projects/noise-da
Open Datasets Yes For image denoising, we follow previous works (Zhang et al., 2018a; Zamir et al., 2022) and construct the synthetic training dataset based on DIV2K (Timofte et al., 2017), Flickr2K (Nah et al., 2019), WED (Ma et al., 2016), and BSD (Martin et al., 2001). ... We use the training dataset of SIDD (Abdelhamed et al., 2018) as the real-world data. For image deraining, the synthetic and real-world training datasets are respectively obtained from Rain13K (Yang et al., 2017) and SPA (Wang et al., 2019). For image deblurring, Go Pro (Nah et al., 2017) and Real Blur-J (Rim et al., 2020) are selected as the synthetic and real-world training datasets, respectively. ... For large-scale unpaired clean images, all images in the MS-COCO dataset (Lin et al., 2014) are used. The test images of the real-world datasets (SIDD, SPA, Real Blur-J) are employed to evaluate the performance of the corresponding image restoration models.
Dataset Splits Yes We use the training dataset of SIDD (Abdelhamed et al., 2018) as the real-world data. ... The test images of the real-world datasets (SIDD, SPA, Real Blur-J) are employed to evaluate the performance of the corresponding image restoration models.
Hardware Specification No No specific hardware details (GPU/CPU models, processor types, memory amounts, or detailed computer specifications) are mentioned in the paper.
Software Dependencies No The paper mentions software components and algorithms (e.g., Adam algorithm, U-Net architecture) but does not provide specific version numbers for any key software dependencies like programming languages or libraries.
Experiment Setup Yes Our model is trained with a fixed learning rate 5e-5 using Adam (Kingma & Ba, 2014) algorithm and the batch size is set to 40. Both the restoration and diffusion networks are trained on 128 128 patches, which are processed with random cropping and rotation for data augmentation. ... To train the diffusion model, we adopt α conditioning and the linear noise schedule ranging from 1e-6 to 1e-2 following previous works (Saharia et al., 2022a;c; Chen et al., 2020). Moreover, the EMA strategy with a decaying factor of 0.9999 is also used across our experiments. ... λDif = 2 / (1 + exp(-γp)) - 1 * β, where γ and β are empirically set to 5 and 0.2 in all experiments, respectively.