Dataset for Simultaneous Enhancement and Super-Resolution (SESR) of underwater imagery.
Introduced in this paper: http://www.roboticsproceedings.org/rss16/p018.pdf (Pre-print)
Code repository: https://github.com/xahidbuffon/Deep-SESR
Project page: image-enhancement-and-super-resolution/deep-sesr
Folder structure:
- train_val/ contains 1500 paired samples for training/validation
- TEST/ contains 120 paired samples for benchmark evaluation
Paired data for each set:
- hr/ contains high-resolution ground truth images: Y
- mask/ contains ground truth saliency maps: S
- lrd/ contains low-resolution distorted images: X
Possible Usage:
- Train SESR models: {X} → {Y}
- Train image enhancement models: {X} → {Y} (downscaled)
- Train image super-resolution models: {Y} (downscaled) → {Y}
- Train saliency prediction models: {Y} (downscaled) → {S}
The Y images are of size 640 x 480. The X and S are of size 320 x 240. Hence, the default configuration is for training 2x models; rescale X and S for other configurations: 3x, 4x, etc.