MMQA

Code for MMQA

Dataset

Dataset	Links
TID2013	http://www.ponomarenko.info/tid2013.htm
KADID10K	http://database.mmsp-kn.de/kadid-10k-database.html
LIVE-C	https://live.ece.utexas.edu/research/ChallengeDB/index.html
LIVE	https://live.ece.utexas.edu/research/Quality/subjective.htm
CSIQ	https://s2.smu.edu/~eclarson/csiq.html
SPAQ	https://github.com/h4nwei/SPAQ
KonIQ	http://database.mmsp-kn.de/koniq-10k-database.html
AGIQA	https://github.com/lcysyzxdxc/AGIQA-3k-Database

Requirements

• PyTorch=1.7.0
• Torchvision=0.8.1
• numpy=1.21.6
• scipy=1.7.3
• h5py=3.7.0
• opencv-python =4.7.0.72

Usages

Screen out salient regions and non-salient regions of images

If you want to get salient regions and non-salient regions, you need to run screen_salient_data.py.

screen_salient_data.py

You will get new datasets, and these datasets are inputs of the model.

Meta training on IQA databases

First you need to modify config paramters to make sure the database path is correct. Meta training our model on IQA Dataset.

MMQA_newload.py

Some available options:

• --dataset: Meta training dataset, support datasets: TID2013 |KADID10K| LIVE | CSIQ | .
• --lr: Learning rate.
• --save_path: Model and paramter save path.
• --batch_size: Batch size.
• --epochs:Epochs

If you want to repartition the dataset, you'll need to make a new mat file instead.

### Fine-tuning for different datasets

FineTune_newload.py

Some available options:
* `--dataset_dir`:  Fine-tuning dataset image path.
* `--model_file`: Model and paramter path.
* `--dataset`:  Testing dataset, support datasets:  LIVE-C | SPAQ | KonIQ | CSIQ| AGIQA |.
* `--predict_save_path`: The plcc and srcc are recorded in TID2013_KADID_LIVEC.txt or ew_load_scores.csv.