Versatile Denoising-Based Approximate Message Passing for Compressive Sensing.

Area of Science:

• Signal Processing
• Image Reconstruction
• Machine Learning

Background:

• Approximate message passing (AMP) algorithms are crucial for compressive sensing (CS) reconstruction.
• The performance of AMP-based CS reconstruction heavily relies on the effectiveness of the denoising operator.
• Existing methods often use generic denoising models, limiting reconstruction performance.

Purpose of the Study:

• To propose a novel versatile denoising-based approximate message passing (VD-AMP) model for improved CS recovery.
• To enhance the denoising capabilities within the AMP framework.
• To address distortion artifacts in CS reconstruction, especially at low sampling rates.

Main Methods:

• Developed a versatile denoising-based approximate message passing (VD-AMP) model.
• Designed a double encoder-decoder denoising network (DEDNet) for Gaussian denoising.
• Implemented a fine-grained noise level division (FNLD) strategy to optimize DEDNet.
• Introduced an anti-aliasing sampling (AS) method using DEDNet to reduce aliasing artifacts.

Main Results:

• The proposed DEDNet demonstrated impressive performance in Gaussian denoising.
• The VD-AMP model significantly improved CS reconstruction performance.
• Achieved up to 2 dB gains in Peak Signal-to-Noise Ratio (PSNR) compared to state-of-the-art methods.
• The anti-aliasing sampling effectively reduced distortion artifacts at low sampling rates.

Conclusions:

• The proposed VD-AMP model, integrating DEDNet and FNLD, offers superior CS reconstruction.
• The anti-aliasing sampling component is crucial for mitigating artifacts in low-rate scenarios.
• VD-AMP represents a significant advancement over existing CS reconstruction techniques.