gokhankocmarli/inline-digital-holography-v3

---
license: mit
task_categories:
- image-to-image
- image-segmentation
- image-classification
- image-feature-extraction
- zero-shot-image-classification
tags:
- holography
- hologram
- inline-holography
- inline
- reconstruction
- synthetic
- signal
pretty_name: Synthetic Inline Holographical Images v3 (224px Highly Diverse)
size_categories:
- 10K<n<100K
---

# Dataset Card for Synthetic Inline Holographical Images v3 (224px Highly Diverse)

This dataset provides synthetic image triplets representing **inline holographical imaging** in a simulated environment. This version (v3) uses a native **224x224 resolution** optimized for modern Vision Transformers (ViT, Swin) and contains 25,000 samples across 8 noise configurations.

Each data sample consists of:
1. An **object-domain field** (ground truth),
2. Its corresponding **forward-propagated hologram** (the inline holographic pattern at the sensor plane, intensity),
3. The **numerically reconstructed image** (via angular spectrum method).

The dataset is intended to facilitate research in **computational imaging, holographic reconstruction, phase retrieval**, and **machine learning-based hologram analysis**.

It is primarily used in conjunction with the open-source project [**Hologen v2**](https://github.com/electricalgorithm/hologen/tree/v2-stable) and [**HoloPASWIN**](https://github.com/electricalgorithm/holopaswin), which provides simulation and learning tools for inline holography.


---

## Dataset Details

The **Synthetic Inline Holographical Images v3** dataset contains triplets of images generated from numerically simulated optical propagation.
The synthetic nature of the data enables large-scale, controllable experiments without the need for physical holographic recording setups.
This makes the dataset especially suitable for deep learning research in holographic imaging, where paired data (object ↔ hologram ↔ reconstruction) are rarely available.

- **Curated and shared by:** Gökhan Koçmarlı (Gyokhan Kochmarla) | [Google Scholar](https://scholar.google.com/citations?hl=en&user=VbZdKZ4AAAAJ) | [GitHub](https://github.com/electricalgorithm/) | [LinkedIn](https://linkedin.com/in/gokhankocmarli/)
- **Funded by [optional]:** Independent research project
- **License:** MIT
- **Repository:** [https://github.com/electricalgorithm/hologen](https://github.com/electricalgorithm/hologen/tree/v2-stable)
- **Data Amount:** 25,000 samples
- **Raw Size:** ~22GB

Dataset consists of 8 noise configuration types with each having 3,125 samples:
- `no_noise`: No noise added.
- `speckle_noise`: Only speckle noise is added.
- `shot_noise`: Only shot noise is added.
- `read_noise`: Only read noise is added.
- `dark_current_noise`: Only dark current noise is added.
- `speckle_shot_noise`: Both speckle and shot noise are added.
- `speckle_shot_read_noise`: The noises speckle, shot and read are added.
- `speckle_shot_read_dark_noise`: All the possible noise types are added.

One can find the configuration name under `config_name` attribute of a sample.

### Simulation Settings

One can reproduce the results using the HoloGen Toolkit with following simulation settings.

| Parameter | Description | Value |
|------------|--------------|-------|
| **Simulation seed** | Random number generator's seed | 10000 |
| **Object height** | Height of both object and sensor plane | 224 pixels |
| **Object width** | Width of both object and sensor plane | 224 pixels |
| **Pixel pitch** | Physical spacing between adjacent pixels | 4.65e-6 meters |
| **Illumination wavelength** | Monochromatic light wavelength | 532e-9 meters |
| **Propagation distance** | Distance between object and sensor planes | 0.02 meters |

| Parameter | Value |
|------------|-------|
| **Speckle Noise's Strength** | 0.15 |
| **Speckle Noise's Roughness** | 1.0 |
| **Read Noise's Sigma** | 10.0 |
| **Dark Noise's Mean** | 20.0 |

---

## Uses

The dataset is designed for:
- Training and evaluating neural networks that reconstruct objects from inline holograms.
- Developing models for **phase retrieval**, **complex field estimation**, and **denoising**.
- Exploring **signal transformation relationships** between object and propagation domains.
- Benchmarking holographic forward and inverse modelling algorithms.

This dataset is **not suitable** for:
- Real-world holography generalisation studies without domain adaptation.
- Tasks requiring physical measurements or phase-accurate calibration data.
- Medical, biometric, or personal data analysis (no human-related content is included).

---

## Dataset Structure

The dataset is stored in **Apache Parquet** format for efficient loading.
- Each sample includes `config_name` attribute to indicate the noise configuration, `global_idx` attribute to track the sample ID across configs, and `sample_idx` attribute to track the sample ID on the same config.
- `ground_truth` and `reconstructed` samples include `real` and `imag` attributes (Complex fields).
- `hologram` samples include `intensity` attribute (Real-valued intensity).

Example Loading Code:
```python
from hologen.dataset import HoloDataset

# Automatically handles loading, normalisation, and splitting
dataset = HoloDataset("path/to/dataset-224", target_size=224, img_dim=224)
hologram, ground_truth = dataset[0]

print(f"Hologram Shape: {hologram.shape}") # (1, 224, 224)
print(f"Ground Truth Shape: {ground_truth.shape}") # (2, 224, 224) -> Real, Imag
```

---

## Dataset Creation

### Curation Rationale

Inline holography involves recording the interference pattern between an object wave and a reference wave. However, collecting large, labelled datasets in laboratory conditions is impractical due to optical setup complexity and noise factors. This dataset provides a **synthetic, physically consistent alternative** that mimics realistic propagation physics using scalar diffraction models.

### Source Data

#### Data Collection and Processing

Images were generated using numerical wave propagation based on the **Angular Spectrum Method (ASM)**, as implemented in the [Hologen](https://github.com/electricalgorithm/hologen) framework. Objects were synthetically generated using shape primitives, textures, and random phase and amplitude patterns. Each object was propagated through a simulated inline holography setup to produce hologram and reconstruction pairs.

#### Who are the source data producers?

- All data were generated algorithmically by Gökhan Koçmarlı using simulation code in *Hologen*.
- No external or third-party datasets were used.

---

### Annotations

No manual annotations are included.
Each triplet is automatically labelled by filename correspondence.

#### Annotation process

Not applicable (fully synthetic, self-labelled data).

#### Who are the annotators?

All data is generated programmatically.

#### Personal and Sensitive Information

This dataset contains **no personal, identifiable, or sensitive information**.
All images are synthetic and algorithmically generated.

---

## Bias, Risks, and Limitations

- As the dataset is fully synthetic, it lacks real-world optical aberrations, noise, and coherence effects that occur in experimental holography.
- Models trained purely on this dataset may require fine-tuning on physical hologram data to generalise effectively.
- The dataset assumes ideal optical parameters (e.g., monochromatic light, planar sensor).

---

### Recommendations

Users should consider:
- Augmenting with noise or real holograms for domain adaptation.
- Interpreting reconstruction metrics (e.g., PSNR, SSIM) relative to synthetic references.
- Avoiding conclusions about physical accuracy without experimental validation.

---

## Article is out! Read HoloPASWIN: Robust Inline Holographic Reconstruction via Physics-Aware Swin Transformers
Please find the article proposing deep learning model that uses the dataset to resolve twin-image and noises for holographical backword propogation:
[https://arxiv.org/abs/2603.04926](https://arxiv.org/abs/2603.04926)

## Citation

**BibTeX:**
```bibtex
@dataset{kochmarla2026synthetic_inline_holographical_images_v3,
author = {Gökhan Koçmarlı},
title = {Synthetic Inline Holographical Images v3 (224px Highly Diverse)},
year = {2026},
url = {https://huggingface.co/datasets/electricalgorithm/inline-digital-holography-v3},
note = {Synthetic dataset for inline holography simulation and reconstruction. Optimized for ViT inputs.}
}
```