A Real-Time and Robust Neural Network Model for Low-Measurement-Rate Compressed-Sensing Image Reconstruction

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Compressed sensing (CS) faces challenges in real-world applications due to high image reconstruction complexity.
• Existing end-to-end learning methods for CS lack theoretical guarantees for robust reconstruction.

Purpose of the Study:

• To propose RootsNet, a novel neural network integrating CS for robust and efficient image reconstruction.
• To address the limitations of traditional CS methods and current deep learning approaches.

Main Methods:

• Integration of the CS mechanism within a neural network architecture (RootsNet) to prevent error propagation.
• Development of a system capable of real-time sensing and reconstruction at extremely low measurement rates.

Main Results:

• RootsNet achieves real-time reconstruction with theoretical guarantees for robustness.
• Successfully reconstructed images at extremely low measurement rates, surpassing traditional optimization-theory-based methods.
• Demonstrated significant improvements in two real-world applications (microwave imaging, pipeline inspection), saving measurement time and data.

Conclusions:

• RootsNet offers superior uncertainty performance, efficiency, and reconstruction quality, especially under super low-measurement rates.
• The proposed method overcomes limitations of existing CS techniques, enabling practical real-world applications.