Self-Supervised Denoising Image Filter Based on Recursive Deep Neural Network Structure.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Traditional noise removal methods often require paired noisy and clean data, limiting their applicability.
• Existing deep learning approaches for noise removal can struggle with performance and detail preservation.

Purpose of the Study:

• To introduce a novel deep neural network-based noise removal method capable of handling various noise types without paired data.
• To enhance denoised image quality through improved detail learning and model performance.

Main Methods:

• Development of the noise-to-blur-estimated clean (N2BeC) model, a deep neural network architecture.
• Implementation of a stage-dependent loss function to regularize existing functions and improve detail learning.
• Incorporation of a recursive learning stage for additional opportunities to learn image details.
• Determination of essential hyperparameters through simulations.

Main Results:

• The N2BeC model demonstrated superior performance, achieving over 1 dB improvement compared to the existing noise-to-blur model.
• The proposed method effectively removes various noise types without requiring paired noisy and clean training data.
• The stage-dependent loss function and recursive learning contributed to better preservation of image details.

Conclusions:

• The proposed N2BeC model offers a significant advancement in unsupervised noise removal using deep neural networks.
• The novel loss function and recursive learning strategy are effective in enhancing denoised image quality and detail preservation.
• This method provides a promising solution for real-world image denoising applications where paired data is unavailable.