QAMD:Quality-Aware Blind Image Motion Deblurring

This is the source code for QAMD:Quality-Aware Blind Image Motion Deblurring.

Dependencies and Installation

Pytorch: 2.1.2

CUDA: 12.1

Python: 3.11

For test:

1. Motion deblurring

The pre-trained model on the GOPRO dataset can be downloaded from: Pre-trained models. Please download the file and put them in the same folder of code, and then create a 'results' folder with subfolders of each dataset (such as 'Hide', 'RealJ'). Finally, run 'deblurring_demo.py' for motion deblurring, and the deblurred images will be saved in the subfolder. The model in 'myrestormer_arch.py' is modified from open accessed source code of Restormer and retrained with patchsize of 128*128 on our device (NVIDIA TITANXP).

2. Calculate values of DISTS and LPIPS

The files in the folder of 'lpips' are obtained from open accessed source code of LPIPS. The files of 'DISTS_pt.py' is modified from open accessed source code of DISTS and 'weights.pt' contains the pre-trained weight. To calculate values of DISTS and LPIPS, please run 'test_crossdatasets_dists_lpips_demo.py' and a 'mat' file containing all values of DISTS and LPIPS will be saved.

For train:

The training code can be available at the 'training' folder.

If you like this work, please cite:

{

author={Tianshu Song, Leida Li, Jinjian Wu, Weisheng Dong, Deqiang Cheng,},

journal={Pattern Recognition},

title={Quality-aware blind image motion deblurring},

volume = {153},

pages = {110568},

year = {2024},

doi = {https://doi.org/10.1016/j.patcog.2024.110568},

}

License

This repository is released under the Apache 2.0 license.