REDDIT · REDDIT// 29d agoRESEARCH PAPER

HoloPASWIN brings physics-aware Swin Transformers to holographic reconstruction

HoloPASWIN is a new research framework for inline digital holography that combines a Swin Transformer U-Net with a differentiable Angular Spectrum physics model to suppress twin-image artifacts. The authors report training on 25,000 noise-augmented samples and claim roughly a 15 dB PSNR gain over standard ASM baselines, with code and paper now public.

// ANALYSIS

This is a credible physics-informed vision paper that matters more for computational imaging than mainstream AI app builders, but it is technically solid and reproducible-minded.

–Blending transformer global context with explicit wave-propagation constraints is a meaningful architecture choice, not just model swapping.
–The noise modeling (shot, read, dark, speckle) makes results more believable than clean-only synthetic demos.
–Open repo + dataset hooks + arXiv release lowers replication friction for labs working on phase retrieval and holography.
–The main caveat is compute cost: differentiable physics in-loop usually improves fidelity at the expense of training efficiency.

// TAGS

holopaswinresearchopen-sourceholographycomputer-visionphysics-informed-learningswin-transformer

DISCOVERED

29d ago

2026-03-14

PUBLISHED

30d ago

2026-03-13

RELEVANCE

7/ 10

AUTHOR

electricalgorithm