Metalens-HyperKvasir

Metalens-HyperKvasir is a paired image restoration dataset derived from Hyper-Kvasir for synthetic metalens image restoration experiments.

Each sample contains:

• gt: the clean target image.
• meta: the corresponding synthetically degraded metalens observation.

The released split is fixed and contains 8,229 paired samples:

Split	Pairs
Train	6,500
Validation	1,729
Total	8,229

The repository contains 16,458 image files plus split_manifest_fixed_sorted.csv. The uploaded release was verified against the local source directory by exact relative-path comparison, file-size comparison for all release files, and SHA256 checks on downloaded samples.

Dataset Structure

The repository root is the dataset root:

Metalens-HyperKvasir/
├── train/
│   ├── gt/        # 6,500 clean target images
│   └── meta/      # 6,500 degraded metalens images
├── val/
│   ├── gt/        # 1,729 clean target images
│   └── meta/      # 1,729 degraded metalens images
└── split_manifest_fixed_sorted.csv

Images in gt and meta are paired by identical filenames.

For example:

train/gt/00000001.png
train/meta/00000001.png

form one clean/degraded training pair.

Fixed Split

The dataset uses a fixed filename-sorted split:

• training: 6,500 pairs
• validation: 1,729 pairs

The split is already materialized in the repository. Users should use the provided train/ and val/ directories directly rather than re-splitting the 8,229 samples.

split_manifest_fixed_sorted.csv is included for split auditing and reproducibility.

Data Generation

The clean source images are derived from the Hyper-Kvasir gastrointestinal endoscopy dataset.

Synthetic metalens degradation was generated using a PSF-aware MetalensTransformer pipeline. The point spread function (PSF) is used during the synthetic degradation generation process.

PSF-unavailable restoration protocol

The restoration task represented by this release is intentionally PSF-unavailable:

• the synthetic degradation generator uses PSF information to create the degraded observations;
• the released restoration input is the degraded RGB image in meta;
• the target is the corresponding clean RGB image in gt;
• the PSF itself is not provided as an input to the restoration model during training or inference.

This distinction is important when comparing restoration methods using this dataset.

Intended Use

This dataset is intended for research on topics including:

• blind or PSF-unavailable metalens image restoration;
• image deblurring and degradation-aware restoration;
• computational imaging;
• endoscopic image restoration;
• spatially varying degradation modeling;
• paired image-to-image restoration.

The dataset is intended as a research benchmark and is not intended for clinical diagnosis, clinical decision-making, or direct patient care.

Download

Hugging Face CLI

Install the Hugging Face Hub client:

pip install -U huggingface_hub

Download the complete repository while preserving the released directory structure:

hf download \
  prpanda123/Metalens-HyperKvasir \
  --repo-type dataset \
  --local-dir ./split_6500_1729

After downloading:

split_6500_1729/
├── train/
│   ├── gt/
│   └── meta/
├── val/
│   ├── gt/
│   └── meta/
└── split_manifest_fixed_sorted.csv

Python

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="prpanda123/Metalens-HyperKvasir",
    repo_type="dataset",
    local_dir="./split_6500_1729",
)

Minimal Pair Loading Example

from pathlib import Path
from PIL import Image

root = Path("./split_6500_1729")

gt_path = root / "train" / "gt" / "00000001.png"
meta_path = root / "train" / "meta" / "00000001.png"

gt = Image.open(gt_path).convert("RGB")
degraded = Image.open(meta_path).convert("RGB")

print("GT size:", gt.size)
print("Degraded size:", degraded.size)

For training, pair images by identical relative filenames under the corresponding gt/ and meta/ directories.

Dataset Size and Integrity

The verified release contains:

Item	Count
train/gt	6,500
train/meta	6,500
val/gt	1,729
val/meta	1,729
Image files	16,458
Split manifest	1

Verified release payload size:

26,463,362,804 bytes

Before publication, the uploaded repository was checked against the local release with:

1. exact train/validation counts;
2. exact gt/meta filename pairing;
3. zero train/validation filename overlap;
4. exact local/remote relative-path agreement;
5. exact byte-size agreement for all 16,459 release files;
6. SHA256 equality on downloaded samples from all four image directories.

Source Dataset: Hyper-Kvasir

This dataset is derived from Hyper-Kvasir, a gastrointestinal endoscopy dataset introduced by Borgli et al.

The official Hyper-Kvasir page requires documents and papers that use or refer to Hyper-Kvasir, or report results based on it, to cite the associated publication.

Please cite:

@article{Borgli2020HyperKvasir,
  title   = {HyperKvasir, a comprehensive multi-class image and video dataset for gastrointestinal endoscopy},
  author  = {Borgli, Hanna and Thambawita, Vajira and Smedsrud, Pia H. and
             Hicks, Steven and Jha, Debesh and Eskeland, Sigrun L. and
             Randel, Kristin Ranheim and Pogorelov, Konstantin and Lux, Mathias and
             Dang-Nguyen, Duc-Tien and Johansen, Dag and Griwodz, Carsten and
             Stensland, Hakon K. and Garcia-Ceja, Enrique and Schmidt, Peter T. and
             Hammer, Hugo L. and Riegler, Michael A. and Halvorsen, Pal and
             de Lange, Thomas},
  journal = {Scientific Data},
  volume  = {7},
  number  = {1},
  pages   = {283},
  year    = {2020},
  doi     = {10.1038/s41597-020-00622-y}
}

Hyper-Kvasir publication:

https://doi.org/10.1038/s41597-020-00622-y

Official Hyper-Kvasir dataset page:

https://datasets.simula.no/hyper-kvasir/

Terms and Attribution

This repository contains derived data based on Hyper-Kvasir.

No additional license is asserted by this Dataset Card for the underlying Hyper-Kvasir source images. Use of this repository must respect the applicable Hyper-Kvasir terms and attribution requirements.

Users should consult the current official Hyper-Kvasir dataset page before redistributing or reusing the underlying image content, and should cite the Hyper-Kvasir publication as required by the upstream dataset terms.

The absence of a license: field in the Dataset Card metadata is intentional: it avoids assigning a license to the derived repository that has not been explicitly established here for the underlying Hyper-Kvasir images.

Limitations

• The degradation is synthetic and may not cover the full distribution of degradations produced by real metalens imaging systems.
• The source images originate from gastrointestinal endoscopy and therefore do not represent general natural-image content.
• Performance on this dataset does not by itself establish performance on real clinical metalens systems.
• The fixed split is provided for reproducibility; results obtained from a different split are not directly comparable to results using the released 6,500/1,729 protocol.
• PSF information is part of the degradation-generation process but is not included as restoration-model input under the intended benchmark protocol.

Repository

Hugging Face dataset repository:

https://huggingface.co/datasets/prpanda123/Metalens-HyperKvasir

A code repository and citation for the associated restoration method can be added here after the public code release and final bibliographic information are available.