gokhankocmarli/inline-digital-holography-v3

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 and 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 | GitHub | LinkedIn
• Funded by [optional]: Independent research project
• License: MIT
• Repository: https://github.com/electricalgorithm/hologen
• 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:

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 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.

---

Citation

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.}
}