Neighbor2Neighbor: A Self-Supervised Framework for Deep Image Denoising.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Deep neural networks have advanced image denoising but require extensive paired noisy-clean data.
• Existing self-supervised methods face challenges like inefficient training, high computational costs, and reliance on noise models.

Purpose of the Study:

• To introduce a self-supervised framework, Neighbor2Neighbor, for deep image denoising that eliminates the need for clean image supervision.
• To theoretically motivate and practically implement a method for training denoising networks using only noisy image pairs.

Main Methods:

• Developed a theoretical framework proving that specific samplers can generate training pairs from noisy images for effective self-supervised denoising.
• Proposed a regularizer to bridge the optimization gap between self-supervised and supervised denoising networks.
• Implemented a simple training scheme using random neighbor sub-samplers and a regularized loss function.
• Introduced BayerEnsemble strategy for adapting Neighbor2Neighbor to raw image denoising.

Main Results:

• The Neighbor2Neighbor framework achieves performance comparable to supervised methods without requiring clean data.
• Demonstrated effectiveness across diverse synthetic and real-world image denoising scenarios.
• The framework leverages advancements in supervised denoising architectures and reduces dependence on noise distribution assumptions.

Conclusions:

• Neighbor2Neighbor offers an efficient and effective self-supervised alternative for deep image denoising.
• The method broadens the applicability of deep learning in image restoration by removing the need for paired data.
• Future work can build upon this framework to further enhance self-supervised image denoising techniques.