Dark Noise Diffusion: Noise Synthesis for Low-Light Image Denoising.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Low-light photography suffers from low signal-to-noise ratios, challenging traditional denoising methods.
• Deep learning for denoising requires large paired datasets, which are difficult to obtain for low-light conditions.
• Synthesizing realistic low-light noise is essential for advancing denoising techniques.

Purpose of the Study:

• To investigate the efficacy of diffusion models in capturing complex low-light noise distributions.
• To develop an adapted diffusion model capable of generating high-precision low-light noise.
• To enable the creation of large synthetic datasets for training low-light denoising networks.

Main Methods:

• Proposed a two-branch diffusion model architecture to differentiate signal-dependent and signal-independent noise.
• Incorporated positional information to accurately model fixed-pattern noise.
• Developed a tailored diffusion noise schedule optimized for low-light noise characteristics.

Main Results:

• Demonstrated that naive diffusion models are insufficient for realistic low-light noise synthesis.
• Achieved high-precision generation of low-light noise distributions through proposed adaptations.
• Generated large synthetic datasets that significantly improved low-light denoising network performance.

Conclusions:

• Adapted diffusion models offer a viable solution for synthesizing realistic low-light noise.
• The proposed model enables state-of-the-art performance in low-light image denoising.
• Further analysis provides deeper insights into low-light noise characteristics and generation.