Disentangling Noise from Images: A Flow-Based Image Denoising Neural Network

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Convolutional Neural Network (CNN)-based methods achieve high accuracy but may introduce artifacts by ignoring clean image distributions.
• Existing denoising techniques often focus on feature extraction, potentially overlooking underlying image properties.

Purpose of the Study:

• To propose a new perspective on image denoising as a distribution learning and disentangling task.
• To develop a framework for denoising based on learning image distributions.
• To present an invertible denoising network (FDN) for enhanced image restoration.

Main Methods:

• Treating image denoising as a distribution learning and disentangling problem.
• Developing a distribution-learning-based denoising framework.
• Introducing an invertible denoising network (FDN) that learns noisy image distributions without assumptions on clean or noise distributions.

Main Results:

• FDN effectively removes synthetic additive white Gaussian noise (AWGN) from various image types.
• The proposed method demonstrates superior performance compared to existing methods in real image denoising.
• FDN achieves better results with fewer parameters and faster processing speeds.

Conclusions:

• Image denoising can be effectively addressed as a distribution learning and disentangling task.
• The proposed invertible denoising network (FDN) offers a promising new direction for high-performance, efficient image denoising.
• FDN surpasses conventional methods in both synthetic and real-world denoising scenarios.