Iterative deep neural networks based on proximal gradient descent for image restoration

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Deep Neural Networks (DNNs) are increasingly used for ill-posed inverse problems.
• Algorithm unfolding networks offer explainability and efficiency for DNNs.

Purpose of the Study:

• To propose a novel end-to-end iterative DNN and its accelerated version for image restoration.
• To enhance feature refinement using the Convolutional Block Attention Module (CBAM) within a U-Net architecture.

Main Methods:

• Developed an algorithm unfolding network framework.
• Incorporated proximal gradient descent from variational models.
• Designed denoiser and reconstruction sub-networks.
• Integrated CBAM into a U-Net for adaptive feature refinement.
• Created an accelerated version with momentum factors.

Main Results:

• Achieved competitive performance in image quality and efficiency for denoising and deblurring tasks.
• Demonstrated superior results compared to several state-of-the-art image restoration networks.

Conclusions:

• The proposed unfolding DNN framework effectively restores images with high quality and efficiency.
• The network architecture is adaptable for various image restoration tasks, including super-resolution and demosaicking.