Dynamic Residual Dense Network for Image Denoising.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep convolutional neural networks excel at image restoration.
• Residual Dense Networks (RDN) show promise in image denoising by utilizing hierarchical features through residual dense blocks (RDBs).
• Existing RDN models struggle with varying noise levels and incur high computational costs with increased depth.

Purpose of the Study:

• To develop a dynamic network capable of handling diverse noise levels efficiently.
• To reduce the computational burden of deep neural networks for image denoising.
• To enhance the adaptability and effectiveness of image denoising for real-world applications.

Main Methods:

• Introduction of the Dynamic Residual Dense Network (DRDN).
• Implementation of a mechanism to selectively skip RDBs based on input image noise levels.
• Inclusion of adjustable denoising strength for manual output optimization.

Main Results:

• The DRDN demonstrates superior performance compared to the standard RDN.
• Achieved a 40-50% reduction in computational cost.
• Outperformed the state-of-the-art CBDNet by 1.34 dB on a real-world noise benchmark.

Conclusions:

• The DRDN offers an effective and computationally efficient solution for image denoising across multiple noise levels.
• Dynamic network architecture allows for adaptive performance and reduced resource utilization.
• The proposed method represents a significant advancement in real-world image restoration.