TeacherIQA

visitors Open issue Closed issue GitHub Stars

This repository is the source code for the paper "Teacher-Guided Learning for Blind Image Quality Assessment".

Framework

Dependencies

  • matplotlib==3.2.2
  • numpy==1.22.3
  • Pillow==9.2.0
  • torch==1.11.0
  • torchvision==0.11.2+cu113

Usages For Testing

You can predict the image quality score for any images with our model which is trained on KonIq-10k dataset.

The pre-trained model can be downloaded from Google drive or Baidu Netdisk (password: b86d).

Please put the pre-trained model into './checkpoint' folder. Then run:

python3 demo.py --input_image ./demoImg/I_02.jpg --pre_train_model ./checkpoint/koniq_teacher_iqa.pkl --crop_times 25

The input image will be randomly crop into 25 patches in size 224 × 224 and the IQA model will predict 25 scores for each patches.

Finally, you will get an average quality score ranging from 0-100. But there exists some cases whose the value may be out of the range. The higher value is, the better image quality is.

Citation

If our work is useful to your research, we will be grateful for you to cite our paper:

@InProceedings{Chen_2022_ACCV,
    author    = {Chen, Zewen and Wang, Juan and Li, Bing and Yuan, Chunfeng and Xiong, Weihua and Cheng, Rui and Hu, Weiming},
    title     = {Teacher-Guided Learning for Blind Image Quality Assessment},
    booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)},
    month     = {December},
    year      = {2022},
    pages     = {2457-2474}
}