NoR-VDPNet++ is a deep-learning based no-reference metric trained on HDR-VDP. Traditionally, HDR-VDP requires a reference image, which is not possible to have in some scenarios.
NoR-VDPNet++ is a no-reference metric, so it requires a single image in order to asses its quality. NoR-VDPNet can be trained on High Dynamic Range (HDR) images or Standard Dynamic Range (SDR) images (i.e., classic 8-bit images).
Requires the PyTorch library along with Image, NumPy, SciPy, Matplotlib, glob2, pandas, and scikit-learn.
As the first step, you need to follow the instructions for installing PyTorch.
To install dependencies, please use the following command:
pip3 install numpy, scipy, matplotlib, glob2, pandas, image, scikit-learn, opencv-python.
To run our metric on a folder of images (i.e., JPEG, PNG, EXR, HDR, and MAT files),
you need to launch the file norvdpnet.py
. Some examples:
Testing SDR images for the trained distortions (see the paper):
python3 norvdpnetpp.py SDR /home/user00/images_to_be_sdr/
Testing HDR images after JPEG-XT compression:
python3 norvdpnetpp.py HDR_COMP /home/user00/images_to_be_hdr/
Testing HDR images after tone mapping operators:
python3 norvdpnetpp.py SDR_TMO /home/user00/images_to_be_sdr/
Testing images after inverse tone mapping operators:
python3 norvdpnetpp.py HDR_ITMO /home/user00/images_to_be_hdr/
Weights can be downloaded at this link.
Note that these weights are meant to model ONLY determined distortions; please see reference to have a complete overview.
There are many people use NoR-VDPNet++ in an appropriate way:
-
Please do not use weights_nor_sdr for HDR images;
-
Please do not use weights_nor_jpg_xt for SDR images;
-
Please do not use weights_nor_tmo for HDR images; only gamma-encoded SDR images!!!
-
Please do not use weights_nor_itmo for SDR images;
-
Please do not use weights for different distortions.
Coming soon.
Coming soon.
If you use NoR-VDPNet in your work, please cite it using this reference:
@ARTICLE{10089442,
author={Banterle, Francesco and Artusi, Alessandro and Moreo, Alejandro and Carrara, Fabio and Cignoni, Paolo},
journal={IEEE Access},
title={NoR-VDPNet++: Real-Time No-Reference Image Quality Metrics},
year={2023},
volume={11},
number={},
pages={34544-34553},
doi={10.1109/ACCESS.2023.3263496}
}