Unsupervised Image Restoration Using Partially Linear Denoisers | JoVE Visualize

Area of Science:

Computer Vision
Machine Learning
Image Processing

Background:

Deep neural networks (DNNs) excel at image restoration but typically require paired noisy and clean images for supervised training.
Acquiring clean ground truth images is often impractical or prohibitively expensive in real-world scenarios.
Existing unsupervised and self-supervised methods offer alternatives but may have limitations in performance.

Purpose of the Study:

To develop a novel deep learning framework for image restoration that can be trained effectively without requiring clean ground truth images.
To demonstrate the method's efficacy in image denoising and its adaptability to other restoration tasks like deblurring.
To achieve competitive or superior performance compared to existing unsupervised and self-supervised deep learning models.

Main Methods:

Proposed a class of structured denoisers decomposable into nonlinear image-dependent mapping, linear noise-dependent term, and a residual term.
Enabled training using only noisy images, provided the noise has zero mean and known variance, without assuming a specific noise distribution.
Extended the structured denoiser framework to address image deblurring challenges.

Main Results:

Demonstrated superior performance in image denoising compared to state-of-the-art unsupervised and self-supervised deep denoising models.
Achieved high-quality results in image deblurring using only a single noisy and blurry observation, approaching the performance of fully supervised methods on benchmark datasets.
Validated the effectiveness of the proposed structured denoiser architecture in unsupervised image restoration.

Conclusions:

The proposed structured denoiser framework offers a viable and effective solution for image restoration tasks where clean data is unavailable.
This method significantly advances unsupervised deep learning for image processing, particularly in denoising and deblurring.
The approach provides a practical alternative to traditional supervised learning, broadening the applicability of deep learning in real-world image restoration.