Blind Universal Bayesian Image Denoising with Gaussian Noise Level Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Blind and universal image denoising aims to remove noise using a single model irrespective of noise levels.
• This approach is practical as it eliminates the need to know noise levels during model development or testing.

Purpose of the Study:

• To propose a theoretically-grounded blind and universal deep learning image denoiser for additive Gaussian noise.
• To evaluate the network's generalization capabilities on unseen noise levels and its performance on real-world images.

Main Methods:

• Developed a deep learning network based on a theoretically derived optimal denoising solution termed 'fusion denoising'.
• Assumed a Gaussian image prior for theoretical derivation.
• Adapted the fusion denoising network architecture for real-world image denoising.

Main Results:

• Synthetic experiments demonstrated the network's strong generalization to unknown additive noise levels.
• The adapted network improved Peak Signal-to-Noise Ratio (PSNR) for real-world grayscale additive image denoising, including unseen noise levels.
• Achieved state-of-the-art performance in color image denoising, improving results across all noise levels by an average of 0.1dB.

Conclusions:

• The proposed fusion denoising network offers a robust solution for blind and universal image denoising.
• The approach effectively handles additive Gaussian noise in both synthetic and real-world scenarios.
• Demonstrated superior performance compared to existing methods for both grayscale and color image denoising.