gokhankocmarli/holopaswin-v3

HoloPASWIN v3

HoloPASWIN (Physics-Aware Swin Transformer for Holographic Reconstruction) is a state-of-the-art model for eliminating twin-image artifacts and recovering complex object fields (amplitude and phase) from single-shot in-line holograms.

Architecture

The model combines the exact physical model of free-space propagation with the powerful representation learning of a Swin Transformer.

1. Physics Module: The input hologram intensity is linearly back-propagated to the object plane using the Angular Spectrum Method (ASM). This creates a "dirty" complex field that contains the focused object but is corrupted by the defocused out-of-focus twin image.
2. Swin Transformer Encoder-Decoder: The dirty field (split into real and imaginary parts) is passed through a hierarchical Swin Transformer U-Net architecture.
3. Residual Correction: Instead of predicting the clean field directly, the network predicts the twin-image artifacts. These are subtracted (via a residual connection) from the dirty input to yield the final, high-fidelity clean reconstruction.

Model Details

• Input: 1-channel Hologram Intensity (B, 1, H, W)
• Output: 2-channel Complex Field (Real, Imaginary) mapped to Amplitude and Phase (B, 2, H, W)
• Encoder Stages: 4 stages with channels [96, 192, 384, 768]
• Parameters: ~30.88M
• Inference Time: ~13.7 ms per image (on modern GPU)

Advancements

HoloPASWIN v3 improves upon traditional CNN-based approaches and iterative solvers:

• Physics-Informed: By letting an exact physical solver handle the propagation mapping, the neural network only has to learn the artifact removal, greatly simplifying the learning objective.
• Global Receptive Field: The Swin Transformer backbone captures long-range spatial dependencies critical in holography, where twin-image interference spreads globally across the entire field of view.
• Single-Shot: Unlike iterative Phase Retrieval algorithms (like Gerchberg-Saxton) that can take hundreds of iterations, this method recovers the field in a single forward pass (~13.7ms).

Outcomes and Results

Evaluated on a test set of 496 holographic samples, HoloPASWIN v3 achieves excellent reconstruction fidelity in both amplitude and phase domains.

Qualitative Results

Below are sample reconstructions showing the Ground-Truth arrays juxtaposed against the Ground-Truth Amplitude & Phase, Zoomed Details, and Error Maps.

Quantitative Metrics

Metric	HoloPASWIN (Ours)	U-Net	ASM (Dirty)	GS (100 iter)
Phase SSIM	0.9862	0.9915	0.3038	0.3038
Phase PSNR (dB)	46.55	46.43	34.69	34.69
Amp SSIM	0.9625	0.9917	0.6128	0.6128
Amp PSNR (dB)	41.98	46.43	30.96	30.96
Complex MSE	0.000968	0.000608	0.0158	0.0158
Inference Time	13.68 ms	23.14 ms	2.40 ms	95.84 ms

Note: While a significantly larger classic U-Net baseline achieves slightly higher raw SSIM metrics on this particular dataset split, HoloPASWIN operates twice as fast (13.7ms vs 23.1ms) while avoiding the common over-smoothing associated with standard CNNs.

How to use

You can load the required classes directly from the repository.

from huggingface_hub import hf_hub_download
import torch

# Load the weights
model_path = hf_hub_download(repo_id="gokhankocmarli/holopaswin-v3", filename="pytorch_model.bin")

# Import the model architecture provided in the same repo
# (Ensure src/holopaswin is in your PYTHONPATH)
from holopaswin.model import HoloPASWIN

# Initialize model
model = HoloPASWIN(
    img_size=224, 
    in_chans=1,
    embed_dim=96,
    depths=[2, 2, 6, 2],
    num_heads=[3, 6, 12, 24],
    window_size=7,
)

# Load state dict
model.load_state_dict(torch.load(model_path))
model.eval()

# Dummy input
dummy_hologram = torch.randn(1, 1, 224, 224)
with torch.no_grad():
    clean_real_imag = model(dummy_hologram)

Citation

If you find this work useful, please cite our paper:

@misc{kocmarli2026holopaswin,
      title={HoloPASWIN: Robust Inline Holographic Reconstruction via Physics-Aware Swin Transformers}, 
      author={Gokhan Kocmarli},
      year={2026},
      eprint={2603.04926},
      archivePrefix={arXiv},
      primaryClass={eess.IV}
}

Or view the paper directly: HoloPASWIN: Robust Inline Holographic Reconstruction via Physics-Aware Swin Transformers