USR-248 Dataset

Dataset for paired training of Underwater Single Image Super-Resolution (SISR) models
Details about the dataset can be found in this paper; please cite it as: 

@article{islam2019srdrm,
    title={Underwater Image Super-Resolution using Deep Residual Multipliers},
    author={Islam, Md Jahidul and Enan, Sadman Sakib and Luo, Peigen and Sattar, Junaed},
    journal={Accepted for publication at the IEEE International Conference on Robotics and Automation (ICRA). arXiv:1909.09437},
    year={2020}
}

Folder structure: 

  • train_val/ contains 1060 paired samples for training/validation
  • TEST/ contains 248 paired samples for benchmark evaluation (do not alter) 

Paired data for each set:    

  • hr/ contains high-resolution ground truth images: 640 x 480
  • Low-resolution pairs are in
    • lr_2x/: 320 x 240
    • lr_3x/: 214 x 160
    • lr_4x/: 160 x 120
    • lr_8x/: 80 x 60
  • Hence, it supports 2x, 3x, 4x, and 8x SISR training