zhangyingyue / image-quality-assessment-by-multiple-models

Using multiple models (CNN, LeNet-5, ResNet, VGG) to implement IQA and compare their performance

Image Quality Assessment by Mutiple Models

Prepare

1. download the database you need and put it under ./data, e.g. ./data/LIVE/fastfading ...

   LIVE (release2 recommended): http://live.ece.utexas.edu/research/Quality/subjective.htm

   tid2008: http://www.ponomarenko.info/tid2008.htm

   tid2013: http://www.ponomarenko.info/tid2013.htm

2. specify datainfo andim_dir in config.yaml

Training

CUDA_VISIBLE_DEVICES=0 python main.py --model='resnet18' --database='LIVE'

• Train/Val/Test split ratio in intra-database experiments can be set in config.yaml (default is 0.6/0.2/0.2).

• Compare different models' performance:

Test Demo

Input an image and output its IQA score

run:

python test_demo.py --im_path=data/images/test_images/blur.jpg --model_file=models/resnet18-LIVE

• --im_path: put your test image in the folder -- data/test_images

• --model_file: choose your trained model, the pre-trained resnet18-LIVE model is given

Visualization

In the server (host:port):

tensorboard --logdir=tensorboard_logs --port=6006

e.g. put the dpai-11 file in logger/test_log, run:

tensorboard --logdir="./logger/test_log" --port=6006

In your PC:

ssh -p port -L 6006:localhost:6006 user@host

• localhost: localhost's IP address
• user: user's name in host
• host: host's IP address

See the visualization in your PC:

Enter localhost:16006 in the browser

Requirements

conda create -n reproducibleresearch pip python=3.6
source activate reproducibleresearch
pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple
source deactive

• Python 3.6.8
• PyTorch 1.3.0
• TensorboardX 1.9, TensorFlow 2.0.0
• pytorch/ignite 0.2.1

Note: You need to install the right CUDA version.