Phase Distribution Alignment (PDA) for Holographic Applications

ACM Multimedia 2025

Seungmi Choi · TaeHwa Lee · Jun Yeong Cha · Suhyun Jo · Hyunmin Ban · Kwan-Jung Oh · Hyunsuk Ko · Hui Yong Kim

Overview

This repository provides the official implementation for the paper:

Phase Distribution Matters: On the Importance of Phase Distribution Alignment (PDA) in Holographic Applications

Phase-only holograms (PoH) encode scene information as a 2π-periodic phase map. Standard image quality metrics (SSIM, PSNR) are not invariant to global phase offsets, producing misleading quality assessments. This repository provides:

• PDA-SSIM — Phase Distribution Alignment SSIM
• PDA-PSNR — Phase Distribution Alignment PSNR
• DPI-DB pipeline — 250 PoH images (50 natural images × 5 CGH methods), 13 distortion types × 5 quality levels
• Full evaluation pipeline — phase-domain metric computation, PLCC/SROCC correlation analysis, and visualization

Key Contributions

1. PDA-SSIM / PDA-PSNR: Phase-aware quality metrics that are invariant to global phase offsets via Distribution-Shifting (PDA) correction. Significantly outperform standard SSIM/PSNR and circular variants in correlation with perceptual NR-domain quality.

2. DPI-DB (Distortion in Phase Image Database): A dedicated benchmark for holographic phase image quality assessment — 250 reference PoH images across 5 CGH methods, 13 distortion types × 5 quality levels (16,250 distorted images total).

Method

PDA-SSIM corrects for global phase offset ambiguity (caused by 2π-periodicity) before computing structural similarity. The key steps:

1. Zero-center: x ← x − 128 (map uint8 [0, 255] → [−128, 127])
2. Double mean removal with symmetric 256-wrap: removes global phase offset
3. β-correction: if mean(|x|) > 64, apply an additional 128-shift to resolve ±π ambiguity
4. Relative alignment: shift the distorted image toward the reference using their mean difference
5. Normalize → radian: x ← (x + 128) / 255, then x ← (1 − x) · 2π − π
6. Decompose: compute cos(x) and sin(x) channels
7. Aggregate: PDA-SSIM = √( (SSIM_cos² + SSIM_sin²) / 2 )

PDA-PSNR applies the same steps 1–5 alignment, then computes PSNR on the aligned radian-domain phase maps.

Installation

conda create -n pda-holo python=3.9
conda activate pda-holo
pip install -r requirements.txt

Notes:

• yuvio is only required for the HEVC/VVC codec pipeline.
• torch / torchvision are only required if running PoH generation (ASM propagation).

Usage

Quick Start — PDA Metrics

import numpy as np
from metrics import Metric_SSIM, Metric_PSNR

# Phase images stored as uint8 [0, 255]  ↔  phase [-π, π]
img_ref  = ...  # ndarray, dtype=uint8 or float64
img_dist = ...

# PDA-SSIM (proposed) — phase-aware
ssim_pda = Metric_SSIM['pda'](img_ref.astype(float), img_dist.astype(float))

# PDA-PSNR (proposed) — phase-aware
rmse, psnr_pda, error_map = Metric_PSNR['pda'](img_ref.astype(float), img_dist.astype(float))

# Comparison baselines
ssim_orig = Metric_SSIM['original'](img_ref.astype(float), img_dist.astype(float), data_range=255)
ssim_circ = Metric_SSIM['circular'](img_ref.astype(float), img_dist.astype(float))

See examples/example_pda_metrics.py for a self-contained runnable demo.

Available Metrics

Registry	Key	Description
Metric_SSIM	'pda'	★ PDA-SSIM — proposed
Metric_SSIM	'original'	Standard SSIM (Wang 2004)
Metric_SSIM	'circular'	C-SSIM
Metric_PSNR	'pda'	★ PDA-PSNR — proposed
Metric_PSNR	'original'	Standard PSNR
Metric_PSNR	'circular'	C-PSNR — (oh2023hevc 2023)

DPI-DB Pipeline

Step 1. Generate PoH images

PoH generation uses ASM wave propagation from Neural Holography (Peng et al., SIGGRAPH Asia 2020).

Step 2. Apply distortions

python make_dataset/add_distortion.py \
    --data_path ./data/poh \
    --methods SGD NHVC DPAC GS

Step 3. Compute phase-domain metrics

python evaluation/cal_phase_metrics.py \
    --data_path ./data/poh \
    --methods SGD NHVC DPAC GS

Output: {method}/graph_pda/{method}.txt

Step 4. Compute NR-domain metrics

First reconstruct distorted PoH images via ASM propagation (see Neural Holography) or the generated batch files from codec/make_bat_eval.py.

python evaluation/cal_nr_metrics.py \
    --data_path ./data/poh \
    --methods SGD NHVC DPAC GS

Output: {method}/{method}.txt

Step 5. Correlation analysis

python evaluation/evaluate.py \
    --data_2d ./data/poh \
    --methods SGD NHVC DPAC GS \
    --out_dir ./results

Output: pcc.txt, data.tsv

Step 6. Draw figures

python evaluation/draw_graph.py \
    --pcc ./results/pcc.txt \
    --data ./results/data.tsv \
    --out_dir ./results

Output: scatter.png, catplot_psnr.png, catplot_ssim.png

Results

PDA-SSIM and PDA-PSNR achieve significantly higher correlation with NR-domain quality scores compared to standard and circular baselines across all CGH methods in DPI-DB.

Project Structure

Phase-Distribution-Alignment/
├── make_dataset/
│   └── add_distortion.py          # 13 distortions × 5 quality levels
│
├── codec/
│   ├── yuv_convert.py             # PNG ↔ YUV444/400 (for HEVC/VVC)
│   ├── make_bat_compress.py       # Generate HEVC/VVC encoding batch files
│   └── make_bat_eval.py           # Generate reconstruction evaluation batch files
│
├── metrics/                       # PDA metrics package
│   ├── __init__.py                # Metric_SSIM, Metric_PSNR registries
│   ├── ssim/
│   │   ├── ssim_original.py       # Standard SSIM (Wang 2004) — 'original'
│   │   ├── ssim_circular.py       # C-SSIM — 256-period circular — 'circular'
│   │   └── ssim_pda.py            # ★ PDA-SSIM (proposed) — 'pda'
│   ├── psnr/
│   │   ├── psnr_original.py       # Standard PSNR — 'original'
│   │   ├── psnr_circular.py       # C-PSNR — 256-period circular — 'circular'
│   │   └── psnr_pda.py            # ★ PDA-PSNR (proposed) — 'pda'
│   └── _utils.py                  # Shared helpers (polar_to_rect, kl)
│
├── evaluation/
│   ├── cal_phase_metrics.py       # Compute phase-domain metrics → {method}/graph_{ver}/
│   ├── cal_nr_metrics.py          # Compute NR-domain metrics → {method}/{method}.txt
│   ├── evaluate.py                # PLCC/SROCC + export pcc.txt / data.tsv
│   └── draw_graph.py              # Scatter + catplots from evaluate.py outputs
│
├── examples/
│   ├── example_pda_metrics.py     # Quick demo: PDA-SSIM / PDA-PSNR
│   └── example_create_dataset.py  # Full DPI-DB pipeline (distort → evaluate → plot)
│
├── requirements.txt
└── README.md

DPI-DB Dataset

DPI-DB (Distortion in Phase Image Database) is a phase-hologram quality assessment database built to evaluate phase-domain metrics against perceptual quality in the numerical reconstruction (NR) domain.

Each sample in DPI-DB consists of four images:

1. Reference natural image — source scene from DIV2K / Unsplash2K
2. Reference PoH — phase-only hologram generated from the natural image via a CGH algorithm
3. Distorted PoH — reference PoH with a distortion applied
4. NR image — numerical reconstruction (via ASM) of the distorted PoH, used as the perceptual quality reference

Reference images

• Natural images: 50 images selected from DIV2K (25) and Unsplash2K (25)
• CGH methods: GS, SGD, DPAC, HoloNet (NHVC), UNet — 5 methods × 50 images = 250 reference PoH images
• Generation parameters: wavelengths 638 / 520 / 450 nm (R/G/B), propagation distance 20 cm, pixel pitch 6.4 µm

Distortion types

13 distortion types applied at 5 quality levels each:

Category	Types
Common noise (7)	White noise (wn), Uniform noise (un), Gaussian blur (gb), Salt & pepper (sp), Global mean shift (ms), Contrast change (cc), Normalization (no)
Block-based (3)	Block Gaussian blur (block_gb), Block global mean shift (block_ms), Block contrast change (block_cc)
Compression (3)	HEVC (hevc), VVC (vvc), HEVC for PoH (hevc2)

In total, DPI-DB contains 16,250 distorted PoH images (250 reference PoH × 13 distortion types × 5 quality levels).

Evaluation protocol

Each distorted phase image is paired with its numerical reconstruction (NR) obtained via Angular Spectrum Method (ASM) propagation. Phase-domain metrics are evaluated by their Pearson Linear Correlation Coefficient (PLCC) and Spearman Rank-Order Correlation Coefficient (SROCC) against NR-domain SSIM and PSNR scores.

🔗 Third-Party Code

The ASM (Angular Spectrum Method) wave propagation used for dataset construction is based on Neural Holography (Peng et al., SIGGRAPH Asia 2020).

GitHub — released under CC BY-NC 4.0

@article{peng2020neural,
  title     = {Neural Holography with Camera-in-the-loop Training},
  author    = {Peng, Yifan and Choi, Suyeon and Padmanaban, Nitish and Wetzstein, Gordon},
  journal   = {ACM Transactions on Graphics (TOG)},
  volume    = {39},
  number    = {6},
  pages     = {1--14},
  year      = {2020},
  publisher = {ACM New York, NY, USA}
}

The circular baseline (C-PSNR / C-SSIM) is based on HEVC extension for phase hologram compression (Oh et al., Optics Express 2023).

@article{oh2023hevc,
  title   = {HEVC extension for phase hologram compression},
  author  = {Oh, Kwan-Jung and Ban, Hyunmin and Choi, Seungmi and Ko, Hyunsuk and Kim, Hui Yong},
  journal = {Optics Express},
  volume  = {31},
  number  = {6},
  pages   = {9146--9164},
  year    = {2023},
  doi     = {10.1364/OE.479281}
}

Citation

@inproceedings{choi2025pda,
  title     = {Phase Distribution Matters: On the Importance of Phase Distribution Alignment (PDA) in Holographic Applications},
  booktitle = {Proceedings of the ACM International Conference on Multimedia},
  year      = {2025},
}

Acknowledgements

This work was supported by Samsung Research Funding & Incubation Center of Samsung Electronics under Project Number SRFC-IT2201-03. And this work was partly supported by Institute of Information & communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT) (No.RS-2022-00155911, Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University)) for setting up some of computing environments needed for our experiments.

Contact

For questions or issues, please open a GitHub issue or contact the authors via vmlab.khu.ac.kr.

License

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike International 4.0 License.