D2S-RSG-SSD: Dual Double-Sampling With Random Sub-Samples Generation for Self-Supervised Real Image Denoising.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Self-supervised image denoising shows promise, particularly with Blind-Spot Networks (BSNs).
• Existing BSNs face limitations: unrealistic noise assumptions, information loss from centroid masking, noise overfitting, and checkerboard artifacts.
• These limitations hinder performance in real-world, complex noise scenarios.

Purpose of the Study:

• To develop a novel self-supervised denoising framework addressing the limitations of current BSNs.
• To improve denoising performance and robustness in real-world image conditions.
• To reduce noise, mitigate information loss, and eliminate sampling-induced artifacts.

Main Methods:

• Introduced Dual Double-Sampling with Random Sub-samples Generation (D2S-RSG-SSD) framework.
• Employed Random Sub-samples Generation (RSG) and a cross-paired loss ($\mathcal {L}_{RSG}$) to break noise dependence.
• Utilized a Dual Double-Sampling (D2S) strategy with a dual-branch architecture and fixed sampling patterns to reduce information loss and noise overfitting.
• Developed a dedicated artifact remover to address checkerboard artifacts, preserving image details.

Main Results:

• The D2S-RSG-SSD framework demonstrated robustness and generalizability on benchmark real-noise datasets and self-captured images.
• Achieved superior performance compared to existing self-supervised denoising methods.
• Effectively reduced noise, minimized information loss, and corrected checkerboard artifacts without over-smoothing.

Conclusions:

• The proposed D2S-RSG-SSD framework offers a significant advancement in self-supervised image denoising.
• The framework successfully overcomes key limitations of previous methods, enabling effective denoising in practical scenarios.
• The approach is compatible with various advanced denoising networks, offering flexibility beyond BSNs.