A Highly Robust Encoder-Decoder Network with Multi-Scale Feature Enhancement and Attention Gate for the Reduction of...

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Image distortions from environmental factors and technical limitations necessitate effective denoising.
• Mixed noise, including salt-and-pepper and Gaussian noise, poses a significant challenge in digital image processing.

Purpose of the Study:

• To introduce a novel Highly Robust Encoder-Decoder Network (HREDN) for effective image denoising.
• To enhance the preservation of image details and structural features during the denoising process.

Main Methods:

• The HREDN incorporates a multi-scale feature enhancement block in the encoder for comprehensive feature capture.
• Skip connections with integrated attention gates are utilized to transfer relevant features and mitigate information loss.
• The network architecture is designed to handle complex noise patterns and preserve structural integrity.

Main Results:

• HREDN demonstrates superior performance in removing mixed salt-and-pepper and Gaussian noise.
• The proposed method excels in preserving edge details and structural features in denoised images.
• Quantitative and qualitative evaluations show HREDN outperforming state-of-the-art denoising techniques.

Conclusions:

• HREDN offers a robust solution for image denoising across diverse domains, including facial, medical, and remote sensing.
• The model effectively handles complex noise scenarios and varying image resolutions.
• The attention-gated skip connections are key to preserving essential image information.