Optimal Transport for Unsupervised Denoising Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Unsupervised denoising methods often rely on signal/degradation assumptions, limiting real-world performance.
• Developing optimal criteria for unsupervised denoising without prior knowledge remains a challenge.

Purpose of the Study:

• To propose a new criterion for unsupervised denoising learning based on optimal transport theory.
• To achieve maximal signal information preservation and perceptual reconstruction without prior degradation model knowledge.

Main Methods:

• Utilized optimal transport theory to formulate a novel criterion for unsupervised denoising.
• Implemented a relaxed unconstrained formulation proven to yield the same theoretical solution as the constrained version.
• Validated the method on synthetic and diverse real-world image datasets (photographic, microscopy, depth).

Main Results:

• The proposed method demonstrates comparable or superior performance to supervised denoising techniques.
• Achieved better perceptual quality and higher peak signal-to-noise ratio (PSNR) than supervised methods, especially for complex noise and microscopy images.
• Showcased remarkable superiority in challenging conditions, such as raw depth images with complex noise.

Conclusions:

• The optimal transport-based criterion offers an effective approach to unsupervised denoising.
• The method achieves a balance between information preservation and perceptual quality, outperforming existing techniques in various scenarios.
• The proposed method shows significant potential for practical applications in image denoising, particularly in adverse conditions.