A Parallel Image Denoising Network Based on Nonparametric Attention and Multiscale Feature Fusion

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Convolutional neural networks (CNNs) excel at image denoising but face challenges with information loss in single-branch models.
• Deep CNNs often exhibit inadequate edge feature extraction and performance saturation.
• Existing methods struggle to balance effective noise removal with the preservation of crucial image details like edges and textures.

Purpose of the Study:

• To propose a novel two-branch convolutional image denoising network.
• To enhance feature extraction capabilities and improve denoising performance.
• To better recover image edge and texture information lost during the denoising process.

Main Methods:

• Employs a two-branch network architecture with complementary structures for deep feature extraction.
• Utilizes densely connected blocks for local feature extraction and dilated convolutions for global feature extraction.
• Integrates a nonparametric attention mechanism for focused feature learning and multiscale feature fusion for robust feature representation.

Main Results:

• The proposed network demonstrated superior objective indexes across multiple standard test datasets (Set12, BSD68, Set5, CBSD68, SIDD).
• Experimental results show the algorithm outperforms several mainstream denoising methods.
• The method effectively retains original image edge and texture information while achieving significant noise reduction.

Conclusions:

• The two-branch network effectively addresses limitations of single-branch models and deep CNN saturation.
• The integration of nonparametric attention and multiscale fusion enhances denoising efficacy and detail preservation.
• This approach offers a promising new direction for deep neural network-based image denoising research.