Noisy-As-Clean: Learning Self-supervised Denoising from Corrupted Image.

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Supervised deep networks excel at image denoising but require paired noisy and clean images.
• Unsupervised methods train on noisy images only, often failing to adapt to specific image priors or noise statistics.
• A domain gap exists between training and testing data, hindering performance on unseen corrupted images, especially with realistic noise.

Purpose of the Study:

• To propose a novel self-supervised learning strategy for image denoising that overcomes the domain gap problem.
• To enable effective denoising using only corrupted images, adapting to their unique characteristics.

Main Methods:

• Introduced the "Noisy-As-Clean" (NAC) strategy for training self-supervised denoising networks.
• NAC uses the corrupted test image as the target and a slightly modified version as input, creating synthetic noisy pairs.
• Demonstrated feasibility of learning optimal parameters with weak noise using only the corrupted image.

Main Results:

• DnCNN and ResNet networks trained with NAC achieved performance comparable to or better than supervised methods.
• NAC-trained networks outperformed previous unsupervised and self-supervised denoising approaches.
• Effective noise removal was shown on both synthetic and realistic noise types.

Conclusions:

• The proposed NAC strategy is a viable and effective self-supervised approach for image denoising.
• NAC mitigates the domain gap problem by adapting to individual image characteristics and noise statistics.
• This method offers a promising alternative for image denoising when clean training data is unavailable.