NesTD-Net: Deep NESTA-Inspired Unfolding Network With Dual-Path Deblocking Structure for Image Compressive Sensing.

Area of Science:

• Signal Processing
• Image Reconstruction
• Deep Learning

Background:

• Deep compressive sensing (CS) is vital for image acquisition and reconstruction.
• Existing deep learning (DL)-based CS methods struggle with block artifacts and information loss, especially at low sampling rates, degrading reconstructed image details.

Purpose of the Study:

• To introduce NesTD-Net, an advanced unfolding-based deep learning architecture for high-quality image CS reconstruction.
• To address limitations of current DL-based CS methods, particularly regarding artifact reduction and detail preservation.

Main Methods:

• Developed NesTD-Net, an unfolding architecture inspired by the NESTA algorithm, integrating DL modules into iterative reconstruction.
• Employed a learned sampling matrix for measurements and an initialization module for initial estimates.
• Incorporated NESTA-derived Iteration Sub-Modules (Yk, Zk, Xk) for iterative l1-norm CS reconstruction.
• Introduced a Dual-Path Deblocking Structure (DPDS) to enhance feature flow and mitigate block artifacts.

Main Results:

• NesTD-Net demonstrated superior performance over state-of-the-art methods in image quality metrics (SSIM, PSNR).
• The method achieved enhanced visual perception and detail reconstruction, particularly at low sampling rates.
• The DPDS module proved versatile and integrable with other unfolding-based methods.

Conclusions:

• NesTD-Net effectively overcomes challenges in deep CS image reconstruction, delivering high-quality results.
• The proposed architecture and DPDS significantly improve detail preservation and reduce artifacts.
• NesTD-Net offers a promising advancement for image CS applications.