Image Quality Assessment Toolbox

🚀 Note: This repository is not currently updated. For the latest metrics, consider exploring: https://github.com/chaofengc/IQA-PyTorch

📧 Feel free to contact: ryanxingql@gmail.com.

0. Archive

• v3: Added MS-SSIM index, BRISQUE, and PIQE; reimplemented PSNR and SSIM in Python; removed Ma et al. and PI due to low computation efficiency; removed FID as it is not an image quality evaluator.
• v2: Unified scripts for all algorithms.
• v1: Initial formal release.

1. Content

metric	class	description	better	range	ref
Peak signal-to-noise ratio (PSNR)	FR	The ratio of the maximum pixel intensity to the power of the distortion.	higher	[0, inf)	[WIKI]
Structural similarity (SSIM) index	FR	Local similarity of luminance, contrast and structure of two image.	higher	(?, 1]	[paper] [WIKI]
Multi-scale structural similarity (MS-SSIM) index	FR	Based on SSIM; combine luminance information at the highest resolution level with structure and contrast information at several down-sampled resolutions, or scales.	higher	(?, 1]	[paper] [code]
Learned perceptual image patch similarity (LPIPS)	FR	Obtain L2 distance between AlexNet/SqueezeNet/VGG activations of reference and distorted images; train a predictor to learn the mapping from the distance to similarity score. Trainable.	lower	[0, ?)	[paper] [official repo]
Blind/referenceless image spatial quality evaluator (BRISQUE)	NR	Model Gaussian distributions of mean subtracted contrast normalized (MSCN) features; obtain 36-dim Gaussian parameters; train an SVM to learn the mapping from feature space to quality score.	lower	[0, ?)	[paper]
Natural image quality evaluator (NIQE)	NR	Mahalanobis distance between two multi-variate Gaussian models of 36-dim features from natural (training) and input sharp patches.	lower	[0, ?)	[paper]
Perception based image quality evaluator (PIQE)	NR	Similar to NIQE; block-wise. PIQE is less computationally efficient than NIQE, but it provides local measures of quality in addition to a global quality score.	lower	[0, 100]	[paper]

Notations:

• FR: Full-reference quality metric.
• NR: No-reference quality metric.

Archived:

metric	class	description	better	range	ref	where
Ma et al. (MA)	NR	Extract features in DCT, wavelet and PCA domains; train a regression forest to learn the mapping from feature space to quality score. Very slow!	higher	[0, 10]	[paper] [official repo]	[v2]
perceptual index (PI)	NR	0.5 * ((10 - MA) + NIQE). Very slow due to MA!	lower	[0, ?)	[paper] [official repo]	[v2]
Fréchet inception distance (FID)	FR	Wasserstein-2 distance between two Gaussian models of InceptionV3 activations (fed with reference and distorted image data-sets, respectively).	lower	[0, ?)	[paper] [cleanfid repo]	[v2]

Subjective quality metric(s):

metric	description	better	range	ref
mean opinion score (MOS)	Image rating under certain standards.	higher	[0, 100]	[BT.500]
degradation/difference/differential MOS (DMOS)	Difference between MOS values of reference and distorted images.	lower	[0, 100]	[ref1] [ref2]

2. Dependency

conda create -n iqa python=3.7 -y && conda activate iqa
python -m pip install pyyaml opencv-python tqdm pandas

# for psnr/ssim
python -m pip install scikit-image==0.18.2

# for ms-ssim/lpips
# test under cuda 10.x
python -m pip install torch==1.6.0+cu101 torchvision==0.7.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html

# for lpips
python -m pip install lpips==0.1.3

For BRISQUE and NIQE, MATLAB >= R2017b is required; for PIQE, MATLAB >= R2018b is required.

If you want to use main.py to run MATLAB scripts, i.e., call MATLAB in Python, you should install MATLAB package in Conda environment. Check here. My solution:

# given linux
cd "matlabroot/extern/engines/python"  # e.g., ~/Matlab/R2019b/extern/engines/python
conda activate iqa && python setup.py install

3. Evaluation

1. Edit opt.yml.
2. Run: conda activate iqa && [CUDA_VISIBLE_DEVICES=0] python main.py -case div2k_qf10 [-opt opt.yml -clean]. [<args>] are optional.
3. Output: CSV log files at ./logs/.

Note:

• tar: target, e.g., enhanced compressed images.
• dst: distorted, e.g., jpeg-compressed images.
• src: source, e.g., raw/pristine images.
• The list of the evaluated images is based on tar_dir.

4. License

We adopt Apache License v2.0. For other licenses, please refer to the references.

If you find this repository helpful, you may cite:

@misc{2021xing3,
  author = {Qunliang Xing},
  title = {Image Quality Assessment Toolbox},
  howpublished = "\url{https://github.com/ryanxingql/image-quality-assessment-toolbox}",
  year = {2021},
  note = "[Online; accessed 11-April-2021]"
}