Physics-Inspired Synthesized Marine Snow Image Dataset and Physics-Inspired Synthesized Underwater Image Dataset
Welcome to Physics-Inspired Synthesized Marine Snow Image Dataset (PHISMID in short) and Physics-Inspired Synthesized Underwater Image Dataset (PHISWID in short).
An example from Physics-Inspired Synthesized Marine Snow Image Dataset. Left: Original underwater image. Right: Synthesized image.
An example from Physics-Inspired Synthesized Underwater Image Dataset. Left: Original underwater image. Right: Synthesized image.
PHISWID is tailored to enhance underwater image processing through physics-inspired image synthesis. PHISWID showcases color degradation and the often-neglected effects of marine snow, a composite of organic matter and sand particles. PHISMID showcases marine snow. We mathematically model the light scattering of marine snow through physics-based underwater image observation model. The modeled artifacts are synthesized with underwater images and construct large-scale pairs of ground-truth and degraded images to calculate objective qualities for underwater image enhancement and to train a deep neural network.
If you use PHISMID or PHISWID in your paper, please cite the following paper. The details for synthesizing marine snow artifacts are also described.
PHISMID: Designed for marine snow removal
PHISWID: Designed for underwater image enhancement/restoration as well as marine snow removal
PHISMID contains 400 image pairs, all having a pixel resolution of 384 x 384. All original underwater images are collected from flickr under Creative Commons Attribution-NonCommercial-ShareAlike 2.0 Generic(CC BY-NC-SA 2.0) License and CC BY 2.0. It consists of an original underwater image and that contains synthesized marine snow artifacts.
PHISWID contains 2264 image pairs, all having a pixel resolution of 384 x 384. All original atmospheric RGB-D images used for PHISWID are collected from NYD-RGB dataset and an outdoor image dataset. An image pair contains one original atmospheric image and one synthesized underwater image degraded by color shift (ueda et al.) and marine snow artifacts.
You can download PHISMID and PHISWID from Google Drive. The file is zipped. After unzipping, you can find original and degraded directories.
The images in original are real underwater images without marine snow or atmospheric images, i.e., ground-truth images. Those in degraded are degraded images with synthesized marine snow artifacts or synthesized color shift and marine snow artifacts.
The images below are examples of PHISMID.
Original underwater image | Synthesized images with marine snow |
---|---|
The images below are examples of PHISWID.
Original underwater image | Synthesized images with color shift and marine snow |
---|---|
The following tables are the current state-of-the-art results for marine snow removal. The average PSNRs/SSIMs are computed over the test datasets. If you would like to update the results, please let us know!!
Method | PSNR | SSIM |
---|---|---|
Median filter (kernel size 3x3) | 30.10 | 0.9907 |
Median filter (kernel size 5x5) | 29.73 | 0.9886 |
Adaptive median filter (kernel size 3x3) | 30.40 | 0.9877 |
Adaptive median filter (kernel size 5x5) | 30.42 | 0.9878 |
U-Net | 37.25 | 0.9930 |
Synthesized image | 30.63 | 0.9873 |
Method | PSNR | SSIM |
---|---|---|
U-Net(UIEB) (C Li et al.) | 20.89 | 0.439 |
U-Net(LSUI) (L Peng et al.) | 21.33 | 0.319 |
U-Net(PHISWID) | 23.97 | 0.714 |
Synthesized image | 19.51 | -0.010 |
The images below are restoration examples for both datasets.
Median filter | Adaptive median filter | U-Net |
---|---|---|
U-Net(UIEB) | U-Net(LSUI) | U-Net(PHISWID) |
---|---|---|
Copyright (c) 2024 Reina Kaneko, Hiroshi Higashi, and Yuichi Tanaka.