Deep nonlocal low-rank regularization for complex-domain pixel super-resolution.

Area of Science:

• Optics and Photonics
• Image Processing
• Computational Imaging

Background:

• Pixel super-resolution (PSR) enhances phase imaging by overcoming sampling limits.
• Existing PSR algorithms struggle with noise sensitivity due to the nonconvexity of phase retrieval and super-resolution.
• Nonlocal low-rank (NLR) regularization offers improved accuracy and robustness in PSR.

Purpose of the Study:

• To develop a more robust and accurate pixel super-resolution (PSR) technique for phase imaging.
• To introduce a novel regularization method inspired by NLR priors for enhanced reconstruction quality.
• To address the limitations of conventional NLR by operating in the deep feature domain.

Main Methods:

• Implemented a plug-and-play framework incorporating nonlocal low-rank (NLR) regularization for PSR.
• Developed the complex-domain nonlocal low-rank network (CNLNet) regularization.
• Performed nonlocal similarity matching and low-rank approximation in the deep feature domain via CNLNet.

Main Results:

• CNLNet-based reconstruction demonstrated superior performance compared to conventional NLR.
• Achieved an average 1.4 dB PSNR improvement over traditional NLR methods.
• Outperformed existing PSR algorithms across various challenging scenarios.

Conclusions:

• The proposed CNLNet regularization offers a state-of-the-art solution for accurate and robust pixel super-resolution in phase imaging.
• Operating in the deep feature domain provides significant advantages over spatial domain methods.
• CNLNet represents a substantial advancement in overcoming noise limitations in phase imaging reconstruction.