SSM-Net: Enhancing Compressed Sensing Image Reconstruction with Mamba Architecture and Fast Iterative Shrinking...

Area of Science:

• Image processing and signal reconstruction
• Machine learning for scientific applications
• Computational imaging

Background:

• Compressed sensing (CS) is vital for high-dimensional imaging but faces challenges in balancing accuracy, efficiency, and convergence.
• Existing CS methods struggle to optimize reconstruction quality with computational speed.

Purpose of the Study:

• To propose SSM-Net, a novel framework addressing limitations in current compressed sensing techniques.
• To enhance reconstruction accuracy, computational efficiency, and convergence speed in high-dimensional image applications.

Main Methods:

• Developed SSM-Net, integrating Mamba's state-space modeling (SSM) with the fast iterative shrinking threshold algorithm (FISTA).
• Incorporated a lightweight sampling module for data compression and an iterative refinement process for deep reconstruction.
• Leveraged SSM's linear complexity for efficient dependency capture and FISTA-inspired momentum for faster convergence.

Main Results:

• SSM-Net demonstrated state-of-the-art reconstruction performance on benchmark datasets.
• Achieved significant reductions in both training and inference reconstruction times compared to existing methods.
• Validated the framework's scalability and practicality for real-time compressed sensing applications.

Conclusions:

• SSM-Net offers a superior balance of reconstruction accuracy, computational efficiency, and convergence speed.
• The Mamba-based approach provides a scalable and practical solution for real-time high-dimensional image reconstruction.
• SSM-Net represents a significant advancement in compressed sensing technology.