Dual-Path Adversarial Denoising Network Based on UNet.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Digital image quality is vital for applications like medical imaging and surveillance.
• Traditional denoising methods often fail to balance noise reduction with detail preservation.
• Adaptability to diverse noise types remains a challenge for existing techniques.

Purpose of the Study:

• To propose a novel three-module architecture for advanced image denoising.
• To enhance the balance between noise removal and fine detail preservation.
• To improve adaptability to various noise types and maintain global structural integrity.

Main Methods:

• A generator creates synthetic noise for data augmentation.
• A dual-path U-Net denoiser with multi-receptive fields preserves details.
• A discriminator provides adversarial feedback for performance enhancement.
• Dual-path adversarial training captures local details and global structures.

Main Results:

• Superior performance demonstrated on SIDD, DND, and PolyU datasets.
• Outperformed state-of-the-art Generative Adversarial Network (GAN) variants.
• Effective noise removal with minimal loss of critical image details was confirmed.

Conclusions:

• The proposed architecture offers robust image denoising for high-fidelity applications.
• It enhances adaptability to complex noise scenarios while preserving structural integrity.
• Provides a versatile tool for image processing tasks requiring detail preservation.