New! IQA-PyTorch implementation

IQA-PyTorch supports UNIQUE now! Can be easily used as follows:

import pyiqa
model = pyiqa.create_metric('unique', as_loss=False)
score = model(img_path)

UNIQUE

The codebase for
Uncertainty-aware blind image quality assessment in the laboratory and wild (TIP2021) and
Learning to blindly assess image quality in the laboratory and wild (ICIP2020)

Prerequisite:

Python 3+
PyTorch 1.4+
Matlab
Successfully tested on Ubuntu18.04, other OS (i.e., other Linux distributions, Windows) should also be ok.

Usage

Sampling image pairs from multiple databases

data_all.m

Combining the sampled pairs to form the training set

combine_train.m

Training on multiple databases for 10 sessions

python Main.py --train True --network basecnn --representation BCNN --ranking True --fidelity True --std_modeling True --std_loss True --margin 0.025 --batch_size 128 --batch_size2 32 --image_size 384 --max_epochs 3 --lr 1e-4 --decay_interval 3 --decay_ratio 0.1 --max_epochs2 12 

(As for ICIP version, set std_loss to False and sample pairs from TID2013 instead of KADID-10K.) (For training with binary labels, set fidelity and std_modeling to False.)

Output predicted quality scores and stds

python Main.py --train False --get_scores True

Result analysis

Compute SRCC/PLCC after nonlinear mapping: result_analysis.m
Compute fidelity loss: eval_fidelity.m

Pre-trained weights

Google Drive: https://drive.google.com/file/d/18oPH4lALm8mSdZh3fWK97MVq9w3BbEua/view?usp=sharing

Baidu: https://pan.baidu.com/s/1KKncQIoQcbxj7fQlSKUBIQ code：yyev

A basic demo that predict the quality of single images.

python demo.py  

Very important ! Make sure that the model has been appropriately set to eval mode !

Link to download the BID dataset

The BID dataset may be difficult to find online, we provide links here:

Google Drive: https://drive.google.com/drive/folders/1Qmtp-Fo1iiQiyf-9uRUpO-YAAM0mcIey?usp=sharing

Baidu: https://pan.baidu.com/s/1TTyb0FJzUdP6muLSbVN3hQ code: ptg0

Training/Testing Data

In addition to the source MATLAB code to generate training/testing data, you may also find the generated files here (If you do not want to generate them yourselves or if you do not have MATLAB):

Google Drive: https://drive.google.com/file/d/1u-6xmedUB0PNA5xM787OY-YfiJg195xA/view

Baidu: https://pan.baidu.com/s/12nb6OTUxnz_rxssg2rthIQ code: 82k3

Citation

@article{zhang2021uncertainty,  
  title   = {Uncertainty-aware blind image quality assessment in the laboratory and wild},  
  author  = {Zhang, Weixia and Ma, Kede and Zhai, Guangtao and Yang, Xiaokang},  
  journal = {IEEE Transactions on Image Processing},    
  volume  = {30},  
  pages   = {3474--3486},  
  month   = {Mar.},  
  year    = {2021}
}

@inproceedings{zhang2020learning,  
  title     = {Learning to blindly assess image quality in the laboratory and wild},  
  author    = {Zhang, Weixia and Ma, Kede and Zhai, Guangtao and Yang, Xiaokang},  
  booktitle = {IEEE International Conference on Image Processing},  
  pages     = {111--115},  
  year      = {2020}
}