Approximating existing visual metrics efficiently using Deep Learning.
This repo contains the code to reproduce the results presented in the following paper:
Artusi, Alessandro, Francesco Banterle, Fabio Carrara, and Alejandro Moreo. "Efficient Evaluation of Image Quality via Deep-Learning Approximation of Perceptual Metrics." IEEE Transactions on Image Processing (2019).
DIQM requires:
- Python 2
- tensorflow 1.2.1
- keras 2.0.5
We strongly suggest using Docker and nvidia-docker to setup the environment and install prerequisites:
docker build -t fabiocarrara/diqm:gpu .
Bring up the environment by issuing the following command in the repo directory:
docker run --runtime nvidia --rm -it -v $PWD:/workdir fabiocarrara/diqm:gpu
Send us a mail and we will be happy to share data and trained models.
The reproduce.sh
file contains all the commands for training the models presented in the paper.
You can reproduce models, predictions, and plots of the paper by issuing:
reproduce.sh
python plot.py runs/
Check plot_p.py
to produce additional paper plots.
Check predict.py
to make predictions using trained models.