Pixel2Pixel: A Pixelwise Approach for Zero-Shot Single Image Denoising.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Image denoising is crucial for enhancing visual quality.
• Traditional methods often require specific noise models or extensive training data.
• Existing deep learning approaches typically rely on large datasets of clean and noisy image pairs.

Purpose of the Study:

• To introduce a novel zero-shot image denoising framework named Pixel2Pixel.
• To enable high-quality image denoising using only the input noisy image.
• To overcome the limitations of data dependency in conventional denoising methods.

Main Methods:

• Leveraging non-local self-similarity within the noisy image to generate training samples.
• Constructing a pixel bank tensor with similar pixels from non-local regions.
• Employing pixel-wise random sampling to create numerous pseudo instances for training.
• Utilizing a compact convolutional neural network architecture.

Main Results:

• Pixel2Pixel successfully generates a large number of training samples from a single noisy image.
• The framework effectively denoises images across various noise types and levels.
• Demonstrated superior performance compared to existing image denoising methods in extensive experiments.
• Showcased strong generalization ability, particularly for real-world noisy scenes.

Conclusions:

• Pixel2Pixel offers an effective zero-shot approach for image denoising.
• The method's reliance on non-local self-similarity eliminates the need for clean training data or noise priors.
• Pixel2Pixel presents a robust and generalizable solution for diverse real-world denoising challenges.