gokhankocmarli/holopaswin-v2

HoloPASWIN v2

A deep learning project for eliminating the twin-image problem in in-line holography using a physics-aware Swin-UNet architecture trained with synthetic holograms generated via the Angular Spectrum Method.

v2 Update (2026): Now operating at native 224x224 resolution with advanced noise robustness (Speckle, Shot, Read, Dark) and optimized Physics-Aware Loss functions.

Find the source on GitHub: HoloPASWIN v2

Training Metadata

Find the training source code using the HoloPASWIN project's GitHub page.

Dataset: gokhankocmarli/inline-digital-holography-v3
Data Configuration: Native 224x224 Parquet Files
Physics: 
- Wavelength: 532 nm
- Pitch: 4.65 μm
- Distance: 20 mm

Samples (Total 25,000):
- Training: 20,000
- Validation: 5,000
- Test (Holdout): 496

Noise Robustness (8 Configs):
- No Noise
- Speckle 
- Shot
- Read
- Dark Current
- (And combinations thereof)

Epochs: 5
Batch Size: 32
Optimizer: AdamW (LR=1e-4)

Accuracy (SOTA)

Evaluated on 496 samples with diverse noise conditions.

============================================================
   FINAL MODEL ACCURACY REPORT (Experiment 9)
============================================================
Evaluated on: 496 samples
------------------------------------------------------------
AMPLITUDE DOMAIN (Absorption/Objects)
  MSE:  0.000130 (±0.000039)
  SSIM: 0.9715   (±0.0042)
  PSNR: 40.62 dB  (±1.25)
------------------------------------------------------------
PHASE DOMAIN (Thickness/Refractive Index)
  MSE:  0.001660 (±0.000578)
  SSIM: 0.9713   (±0.0059)
  PSNR: 44.00 dB  (±1.43)
------------------------------------------------------------
COMPLEX DOMAIN (Overall Fidelity)
  MSE:  0.001675 (±0.000540)
============================================================