Extra Proximal-Gradient Network with Learned Regularization for Image Compressive Sensing Reconstruction.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Applied Mathematics

Background:

• Learned optimization algorithms combine numerical optimization and deep neural networks for inverse problems.
• Image reconstruction is a key application area for solving inverse problems.

Purpose of the Study:

• To propose a novel deep neural network architecture for solving general inverse problems.
• To focus on improving image reconstruction quality using learned regularization.

Main Methods:

• Developed a deep neural network architecture inspired by extra proximal gradient algorithms.
• Incorporated learned regularization with adaptive sparsification, robust shrinkage, and nonlocal operators.

Main Results:

• The proposed network demonstrated improved efficiency and accuracy compared to existing methods.
• Numerical results validated the network's performance on various test problems.

Conclusions:

• The novel deep learning approach effectively addresses inverse problems, particularly in image reconstruction.
• Learned regularization techniques significantly enhance solution quality and method performance.