Holopaswin Reconstruction Eval

holopaswin-reconstruction-eval

HoloPASWIN: Robust Inline Holographic Reconstruction via Physics-Aware Swin Transformers — Kocmarli et al. (2026) (arXiv:2603.04926, 2026)

What this evaluates

Evaluates a deep learning model's ability to reconstruct complex object fields from in-line digital holograms, specifically testing its capacity to suppress twin-image artifacts and maintain reconstruction fidelity under various noise conditions.

Datasets

• Synthetic Inline Holography Dataset — total 25496; splits: train (20000), val (5000), test (496); repo https://github.com/electricalgorithm/holopaswin

Metrics

• reconstruction fidelity (primary) — range: other
  • Not explicitly defined in the provided section; generally refers to quantitative comparison between the predicted complex object field and the ground-truth object field.
• twin-image suppression — range: other
  • Not explicitly defined in the provided section; evaluates the model's ability to eliminate conjugate-free twin artifacts inherent to phase-loss in intensity recording.

Input / output format

Input: 224×224 pixel intensity hologram, normalized by dividing raw 12-bit intensity values by 1000.0.

Output: 224×224 pixel complex object field (real and imaginary components), left unnormalized.

Scoring recipe

pred_field = model.predict(normalized_hologram)
gt_field = load_ground_truth_object_field()
fidelity_score = compute_fidelity_metric(pred_field, gt_field)
twin_suppression_score = compute_twin_suppression_metric(pred_field, gt_field)
return fidelity_score, twin_suppression_score

Common pitfalls

• The dataset is entirely synthetic; the authors explicitly state that validation on experimental data is left for future work.
• The simulation only models objects at a single depth plane (z=0) and does not account for volumetric scattering.
• Twin-image artifacts arise purely from phase-loss during intensity recording, not from out-of-focus objects, which is a common misconception in holography.

Evidence (verbatim from paper)

outperforming CNN-based methods in both twin-image suppression and reconstruction fidelity. A separate external test set of 496 samples was used for final evaluation. The supervised loss and physics consistency loss together guide the network to produce a physically valid solution that eliminates the twin artifact.

Citation

@misc{kochmarli2026holopaswin,
  title={HoloPASWIN: Robust Inline Holographic Reconstruction via Physics-Aware Swin Transformers},
  author={Kocmarli et al. (2026)},
  year={2026},
  note={arXiv:2603.04926}
}

• arXiv: 2603.04926