Stimulating the Diffusion Model for Image Denoising via Adaptive Embedding and Ensembling

Tong Li, Hansen Feng, Lizhi Wang, Zhiwei Xiong, Hua Huang

Paper | Code | Note | Note

• We present a novel strategy called the Diffusion Model for Image Denoising (DMID) by understanding and rethinking the diffusion model from a denoising perspective.
• Our DMID strategy includes an adaptive embedding method that embeds the noisy image into a pre-trained unconditional diffusion model and an adaptive ensembling method that reduces distortion in the denoised image.

Positive2Negative: Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising

Tong Li, Lizhi Wang, Zhiyuan Xu, Lin Zhu, Wanxuan Lu, Hua Huang

Paper | Code

• We propose a new self-supervised single image denoising paradigm, Positive2Negative, to break the information-lossy barrier.
• We propose a data construction method, which constructs multi-scale similar noisy images for training.
• We propose a denoising supervision method, which is theoretically guaranteed to learn robust denoising.

PDE: Gene Effect Inspired Parameter Dynamic Evolution for Low-light Image Enhancement

Tong Li, Lizhi Wang, Hansen Feng, Lin Zhu, Hua Huang

Paper | Code

• We identify and illustrate a counterintuitive phenomenon in existing models, which we refer to as the “gene effect”.
• We propose the PDE method to mitigate the gene effect, primarily relying on the POG technique.
• Experiments show our method mitigate the gene effect while improving the performance of LLIE.

Positive2Negative 期刊拓展

Tong Li, Lizhi Wang, Zhiyuan Xu, Lin Zhu, Wanxuan Lu, Hua Huang

Paper | Code

• We propose a new self-supervised single image denoising paradigm, Positive2Negative, to break the information-lossy barrier. +
• We propose a data construction method, which constructs multi-scale similar noisy images for training. +
• (Private)
• We propose a denoising supervision method, which is theoretically guaranteed to learn robust denoising.

YOND: Practical Blind Raw Image Denoising Free from Camera-Specific Data Dependency

Hansen Feng, Lizhi Wang, Yiqi Huang, Tong Li, Lin Zhu, Hua Huang

Paper | Code

• We introduce a novel blind raw image denoising method. With our method, an AWGN denoiser can generalize to various real raw data with a single training on synthetic datasets. We name our method YOND, as you need nothing else under our method, You Only Need a Denoiser.
• YOND consists of three key modules: the coarse-to-fine noise estimation (CNE), the expectation-matched variance-stabilizing transform (EM-VST), and the SNR-guided denoiser (SNR-Net).
• Extensive experiments across diverse camera datasets, along with flexible solutions for challenging cases, demonstrate the practicality of YOND.