DeepSN-Net: Deep Semi-Smooth Newton Driven Network for Blind Image Restoration

Area of Science:

• Computer Vision
• Machine Learning
• Optimization Algorithms

Background:

• Deep unfolding networks are promising for image restoration but often use slow first-order optimization.
• Existing methods face challenges with convergence speed and learning efficiency.

Purpose of the Study:

• To develop a more efficient and accurate deep unfolding network for image restoration.
• To introduce a second-order optimization approach to address the limitations of current methods.

Main Methods:

• Formulated an improved second-order semi-smooth Newton (ISN) algorithm.
• Developed DeepSN-Net, a novel network architecture based on the ISN algorithm for blind image restoration.
• Designed a unified framework applicable to various degradation conditions and contexts.

Main Results:

• DeepSN-Net demonstrates high learning efficiency and superior restoration accuracy.
• The network exhibits good generalization ability across 11 datasets and three restoration tasks.
• Achieved the first successful implementation of a second-order deep unfolding network for image restoration.

Conclusions:

• DeepSN-Net represents a significant advancement in image restoration by leveraging second-order optimization.
• The proposed method offers a unified, interpretable, and efficient framework for diverse image restoration challenges.
• This work paves the way for future research utilizing second-order optimization in deep unfolding networks.