LSDIR: A Large Scale Dataset for Image Restoration

How to use the dataset?

1. Download the dataset from the LSDIR website or Hugging Face.
2. Load the JSON file for the specific task.
   • train.json: for image restoration tasks based on synthetic data such as image denoising, real-world image super-resolution based on pure data synthesis, JPEG compression artifact removal, Gaussian image deblurring, image demosaicking.
     • path: path to ground-truth images
   • train_X2.json: for X2 image super-resolution
     • path_gt: path to ground-truth images
     • path_lq: path to X2 low-resolution images
   • train_X3.json: for X3 image super-resolution
     • path_gt: path to ground-truth images
     • path_lq: path to X3 low-resolution images
   • train_X4.json: for X4 image super-resolution
     • path_gt: path to ground-truth images
     • path_lq: path to X4 low-resolution images
3. Write your own dataloader based on the loaded JSON file. (We will provide our dataloader soon.)

Stay tuned!

The paper, code, and benchmark will be released soon.