MID: A Self-Supervised Multimodal Iterative Denoising Framework.

Area of Science:

• Computer Vision
• Biomedicine
• Bioinformatics
• Signal Processing

Background:

• Denoising is crucial for vision, medical, and biological applications.
• Real-world data often suffers from complex nonlinear noise.
• Clean target data is frequently unavailable for training.

Purpose of the Study:

• To present MID, a self-supervised iterative denoising framework adaptable across data modalities.
• To enable effective denoising using only noisy observations, addressing the lack of clean targets.

Main Methods:

• MID models observations as intermediate states in a controllable corruption process.
• It employs two networks: a step predictor for corruption stage estimation and a residual predictor for noise removal.
• A first-order local approximation handles nonlinear corruption by enabling iterative restoration in a locally linear regime.

Main Results:

• MID demonstrates robustness and broad applicability across diverse tasks.
• Experiments in computer vision, biomedicine, and bioinformatics show competitive performance against existing methods.
• The framework successfully denoises images, signals, point sets, and sequences.

Conclusions:

• MID offers a versatile and effective self-supervised approach to denoising.
• The framework's ability to handle nonlinear noise and work across modalities makes it highly valuable.
• MID presents a significant advancement in self-supervised learning for data restoration.