ADMM-CSNet: A Deep Learning Approach for Image Compressive Sensing.

Area of Science:

• * Signal Processing
• * Machine Learning
• * Medical Imaging

Background:

• * Compressive sensing (CS) enables image reconstruction from limited sampled data, with applications in medical imaging and remote sensing.
• * Traditional CS methods often require complex optimization and may struggle with diverse imaging scenarios.
• * Deep learning offers data-driven approaches but can sometimes lack interpretability or theoretical grounding.

Purpose of the Study:

• * To propose a novel deep learning architecture, ADMM-CSNet, for enhanced image reconstruction using compressive sensing.
• * To integrate traditional model-based CS principles with data-driven deep learning for improved performance.
• * To develop efficient solvers for a generalized CS model and adapt them into deep learning architectures.

Main Methods:

• * Developed two versions of ADMM-CSNet by combining model-based CS and deep learning.
• * Utilized the Alternating Direction Method of Multipliers (ADMM) algorithm for model optimization.
• * Unrolled and generalized the ADMM algorithm into deep architectures trained end-to-end to learn all parameters.

Main Results:

• * ADMM-CSNet achieved favorable reconstruction accuracy in both complex-valued MR imaging and real-valued natural image reconstruction.
• * The proposed method demonstrated faster computational speed compared to traditional and other deep learning techniques.
• * Successfully learned all parameters of the CS model and ADMM algorithm through discriminative end-to-end training.

Conclusions:

• * ADMM-CSNet offers a powerful hybrid approach for compressive sensing image reconstruction.
• * The architecture provides a balance of accuracy and computational efficiency for various imaging applications.
• * This deep learning framework effectively addresses limitations of traditional CS methods and purely data-driven models.