A Residual UNet Denoising Network Based on Multi-Scale Feature Extraction and Attention-Guided Filter

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Effective image denoising is crucial for high-quality image analysis.
• Traditional methods often struggle to balance noise suppression with detail preservation.
• Deep learning models, particularly Convolutional Neural Networks (CNNs), have shown promise in image restoration tasks.

Purpose of the Study:

• To propose a novel residual UNet denoising network for superior image noise removal.
• To enhance image detail retention and sharpness during the denoising process.
• To introduce innovative components like attention-guided filters and multi-scale feature extraction.

Main Methods:

• A residual UNet architecture was developed, incorporating attention-guided filter blocks and multi-scale feature extraction blocks.
• Multi-scale feature extraction blocks were used as input blocks to broaden the receptive field and capture richer features.
• Attention-guided filter blocks were designed to preserve critical edge information.
• A global residual network strategy was employed to model residual noise, rather than directly predicting clean images.

Main Results:

• The proposed network demonstrated superior performance compared to several state-of-the-art denoising models.
• Experimental results confirmed more effective noise suppression capabilities.
• Significant improvements in image sharpness were observed, indicating better detail preservation.

Conclusions:

• The developed residual UNet denoising network effectively suppresses noise while preserving image details and enhancing sharpness.
• The integration of attention-guided filters and multi-scale feature extraction contributes to the model's improved performance.
• This approach offers a promising solution for high-quality image denoising applications.