SGDNet: An End-to-End Saliency-Guided Deep Neural Network for No-Reference Image Quality Assessment

This repository contains the reference code for our ACM MM 2019 paper. The pdf can be found in this link.

If you use any part of our code, or SGDNet is useful for your research, please consider citing:

@inproceedings{yang2019sgdnet,
  title={SGDNet: An End-to-End Saliency-Guided Deep Neural Network for No-Reference Image Quality Assessment},
  author={Yang, Sheng and Jiang, Qiuping and Lin, Weisi and Wang, Yongtao},
  booktitle={Proceedings of the 27th ACM International Conference on Multimedia},
  year={2019},
  organization={ACM}
}

• Python 2.7
• Keras 2.1.2
• Tensorflow-gpu 1.3.0

• Clone this repo:

git clone https://github.com/ysyscool/SGDNet
cd SGDNet
mv SGDNet/acmmm_release/ SGDNet/
mkdir ../checkpoint/

• Download weights from Google Drive. Put the weights into

cd ../checkpoint/

1. Download the IQA datasets. Their saliency maps, used in our experiments, can be downloaded in this link.
2. Modify the paths in config.yaml And then using the following command to train the model (use knoiq10k and DINet as example)

 CUDA_VISIBLE_DEVICES=0 python main.py  --database=Koniq10k --lr=1e-4 --batch_size=19 --out2dim=1024  --saliency=output --phase=train

For testing, modify the variables of arg (in line 276) as the trained checkpoint name in the main.py. And then using the following command to test the model

CUDA_VISIBLE_DEVICES=0 python main.py  --database=Koniq10k --out2dim=1024 --saliency=output --phase=test

Code and data prepration largely benefits from CNNIQAplusplus by Dingquan Li.