Rotation Equivariant Proximal Operator for Deep Unfolding Methods in Image Restoration

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep unfolding methods combine traditional image processing with deep learning for high interpretability.
• Current methods use CNNs for proximal networks, excelling at translational symmetry but lacking rotation symmetry handling.

Purpose of the Study:

• To develop a rotation equivariant proximal network for deep unfolding frameworks.
• To embed rotation symmetry priors into deep unfolding for improved performance.

Main Methods:

• Designed a novel rotation equivariant proximal network.
• Deduced theoretical equivariant error for arbitrary layers and rotation degrees.
• Integrated the network into existing deep unfolding architectures.

Main Results:

• The proposed network effectively embeds rotation symmetry priors.
• Theoretical analysis provides refined error evaluation for interpretability.
• Experimental validation on super-resolution, medical imaging, and de-raining shows performance enhancement.

Conclusions:

• The rotation equivariant proximal network can replace standard CNN-based networks.
• The method readily enhances state-of-the-art performance in various vision tasks.
• Demonstrates potential for general applicability in computer vision.