Deep Variation Prior: Joint Image Denoising and Noise Variance Estimation Without Clean Data.

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Supervised deep learning excels at image denoising but requires extensive paired noisy and clean data.
• Unsupervised methods offer an alternative but struggle with accurate noise modeling, especially for unknown noise distributions.
• Existing approaches often require separate noise estimation steps, complicating the denoising process.

Purpose of the Study:

• To develop an unsupervised deep learning framework for joint image denoising and noise variance estimation.
• To address the ill-posed nature of learning with unknown noise distributions.
• To eliminate the need for clean training data and external noise estimation procedures.

Main Methods:

• Introduced the Deep Variation Prior (DVP) criterion for evaluating denoisers based on noise variation smoothness.
• Developed an unsupervised deep learning framework that jointly optimizes denoiser and noise variance estimation.
• Leveraged DVP and zero-mean, pixel-wise independent noise assumptions to approximate minimum mean squared error denoisers.

Main Results:

• Achieved image denoising performance comparable to supervised learning methods.
• Demonstrated accurate estimation of noise variances within the joint framework.
• Successfully bypassed the requirement for clean training images and separate noise estimation.

Conclusions:

• The proposed joint unsupervised framework effectively tackles image denoising and noise variance estimation.
• Deep Variation Prior provides a robust criterion for unsupervised denoiser learning.
• This approach offers a powerful alternative for image denoising when clean data is unavailable.