A Scalable Training Strategy for Blind Multi-Distribution Noise Removal.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Image Processing

Background:

• Developing universal denoising networks is challenging due to the trade-off between specialization and generalization.
• The curse of dimensionality complicates training by requiring exponential increases in data for diverse noise specifications.

Purpose of the Study:

• To develop an adaptive-sampling/active-learning strategy for training universal denoising networks.
• To improve upon existing universal denoiser training strategies by extending to higher dimensions and incorporating polynomial approximations.

Main Methods:

• Implemented an adaptive-sampling/active-learning strategy for training denoising networks.
• Incorporated a polynomial approximation of the specification-loss landscape to reduce training time.
• Tested the method on simulated and real-world joint Poisson-Gaussian-Speckle noise.

Main Results:

• Achieved significant reductions in training time (almost two orders of magnitude).
• A single generalist denoiser network demonstrated performance comparable to specialized networks across various noise conditions.
• The adaptive-sampling strategy outperformed uniform sampling in real-world image denoising tasks.

Conclusions:

• The proposed adaptive-sampling strategy enables the creation of effective universal denoising networks.
• This approach overcomes the curse of dimensionality and reduces training complexity.
• The trained generalist denoiser shows robust performance across a wide range of noise types and conditions.