Benchmark study of deep super-resolution models for digital holography: quantitative phase and intensity evaluation

Area of Science:

• Optics and Photonics
• Biomedical Imaging
• Computational Imaging

Background:

• Holographic microscopy offers good axial resolution but is limited in lateral resolution by optical factors, pixel size, and noise.
• These limitations impede the precise reconstruction of fine structures and phase details crucial for biological and material science applications.

Purpose of the Study:

• To evaluate the effectiveness of deep learning super-resolution models in enhancing lateral resolution for digital holographic microscopy.
• To compare the performance of RCAN, SwinIR, and a conditional diffusion network against bicubic spline interpolation.

Main Methods:

• Utilized 1,440 off-axis digital holograms of microbeads, downsampled by 2×, 3×, and 4×.
• Assessed three deep learning models: RCAN, SwinIR, and a conditional diffusion network.
• Performance was quantified using Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), Mean Squared Error (MSE), and phase-derived depth errors.

Main Results:

• RCAN and SwinIR demonstrated superior performance, achieving the most accurate reconstructions.
• These models effectively preserved both structural details and quantitative phase information.
• Deep learning models significantly outperformed bicubic spline interpolation in super-resolution tasks.

Conclusions:

• RCAN and SwinIR are highly effective for super-resolution in holographic microscopy, particularly for preserving phase information.
• The study provides valuable guidance for selecting appropriate deep learning models for phase-focused holographic applications.
• Deep learning offers a promising avenue for overcoming lateral resolution constraints in digital holography.