Structural Similarity-Inspired Unfolding for Lightweight Image Super-Resolution.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Current data-driven image super-resolution (SR) methods increase model complexity to capture context.
• Model-driven SR methods offer compactness but can be limited in performance.
• A hybrid approach could leverage the strengths of both methodologies.

Purpose of the Study:

• To develop an efficient image super-resolution (SR) method by combining data-driven and model-driven approaches.
• To improve SR performance while maintaining model compactness.
• To introduce a novel unfolding-based SR framework inspired by structural similarity.

Main Methods:

• Proposed a Structural Similarity-Inspired Unfolding (SSIU) method for efficient image SR.
• Unfolded an SR optimization function constrained by structural similarity.
• Incorporated Mixed-Scale Gating Modules (MSGM) for feature constraints and Efficient Sparse Attention Modules (ESAM) for sparse activation within each iteration.
• Utilized a Mixture-of-Experts-based Feature Selector (MoE-FS) for multi-level feature integration.

Main Results:

• The SSIU method demonstrated superior performance compared to existing state-of-the-art SR models.
• Achieved significant reductions in parameter count and memory consumption.
• Validated the efficacy and efficiency of the proposed unfolding-inspired network through extensive experiments.

Conclusions:

• The SSIU method effectively combines data-driven and model-driven strategies for efficient image SR.
• The proposed architecture achieves high performance with improved model compactness.
• This work offers a promising direction for developing efficient and effective super-resolution techniques.