Noise2Kernel: Adaptive Self-Supervised Blind Denoising Using a Dilated Convolutional Kernel Architecture.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Unsupervised learning enables deep network training for image denoising without paired noisy and clean images.
• Current methods often assume zero-mean, signal-independent noise, leading to artifacts with unconventional noise statistics.
• Existing blind denoising techniques typically rely on random masking for training invariance.

Purpose of the Study:

• To develop an efficient unsupervised deep network for image denoising.
• To address limitations of current methods regarding unconventional noise statistics and training requirements.
• To improve image denoising performance, especially for noise types lacking prior statistical knowledge.

Main Methods:

• Proposed a dilated convolutional network with an invariant property for efficient kernel-based training, eliminating the need for random masking.
• Introduced an adaptive self-supervision loss function to enhance tolerance for unconventional noise.
• Evaluated the method's effectiveness on various noise types, including salt-and-pepper and hybrid noise.

Main Results:

• The proposed dilated convolutional network enables efficient training without random masking.
• The adaptive self-supervision loss effectively handles unconventional noise statistics, reducing artifacts.
• Demonstrated superior performance compared to state-of-the-art denoising methods on diverse image examples.

Conclusions:

• The developed method offers an effective unsupervised approach for image denoising, particularly for challenging noise conditions.
• The invariant network property and adaptive loss function contribute to robust and artifact-free denoising.
• This work advances unsupervised image denoising by accommodating a wider range of noise models.