PG-SWIQA: A Physics-Guided Spatial-Wavelet Interaction Network for No-Reference Underwater Image Quality Assessment

Requirements

• python=3.8
• torch 2.0.0+cu117 torchvision0.15.1+cu117 torchaudio==2.0.1
• scikit-learn
• pandas
• tensorboardX
• tensorboard
• opencv-python
• imgaug
• timm
• openpyxl
• PyWavelets

File Structures of the Dataset

• Simply place the images in the dataset in the corresponding folder, the labels are already in "mos.xlsx". The folder structure is as follows.

PIGUIQA/
│   ...
├───Data/
│   ├───SAUD2.0/
│   │   ├───mos_result/
│   │   │   ├───mos.xlsx
│   │   │   ├───record.txt
│   │   │   └───results.xlsx
│   │   ├───train/
│   │   │   ├───train_dataset.pth
│   │   │   └───...
│   │   ├───test/
│   │   │   ├───test_dataset.pth
│   │   │   └───...
│   │   ├───001_BL-TM.png
│   │   ├───001_GL-net.png
│   │   └───...
│   ├───SOTA20000/
│   │   └───...
│   ├───UID2021/
│   │   └───...
│   └───UWIQA/
│       └───...
│   ...

• The "train" folder, "test" folder, "record.txt", and "results.xlsx" will be automatically created after running "main.py".

Execution

• Please run "main.py".
• For training, please set "train = True", and set your "data_path". The file structures of the SAUD2.0, SOTA20000, UID2021 and UWIQA have been given. You can also use your own dataset.
• For testing, please set "train = False", and set your "data_path" and "pretrained_model_path".

Record and Result

• The record of the training process and the testing results can be found in "record.txt", and "results.xlsx".