khrynczenko / compimg Goto Github PK

On SSIM wikipedia page it is mentioned that SSIM can be calculated also on Gaussian window thus it would be nice to have it in the package.

In most places I use term similarity (metric) for the methods implemented within the package. I use this term within the docs, code and readme. I think it would be better to use Image Quality Assessment (IQA) term and its different categories full-reference (which is what I call similarity), reduced-reference and no-reference. I think it would be more clear for the users what compimg is about.
Also it would open space for implementing reduced-reference and no-reference methods in the future which are obviously not similarity metrics.

SSIM and GSSIM take very long time to compute even for small images. Most of the time is spent
in convolve function when performing window multiplication and sum. Options are either to optimize current code or to ditch the 'only numpy as dependency' rule and use scipy.

Seems to be reasonable addition. Figure out whether its addition would be valuable and worth the effort.

Running tests in automated manner is should be provided. CircleCI seems popular.

Candidates are:
Pattern Intensity
Normalized Correlation
VSNR
VIF
IFC
VMAF (Netflix)
https://en.wikipedia.org/wiki/Video_quality#Classification_of_objective_video_quality_metrics
https://www.its.bldrdoc.gov/resources/video-quality-research/software.aspx
https://itk.org/Doxygen/html/ImageSimilarityMetricsPage.html
http://library.utia.cas.cz/separaty/2013/RO/kudelka-image%20quality%20assessment.pdf

SSIMplified seems to be interesting read. Mayhap there can be found some application to compimg.

Gradient-Based Structural Similarity Index (GSSIM) according to paper "GRADIENT-BASED STRUCTURAL SIMILARITY FOR IMAGE QUALITY ASSESSMENT".

Although not widely used and of poor performance those metric could be easily added to the package.

Universal Image Quality Index (UIQI) is another full-reference method for quality assessment. https://ece.uwaterloo.ca/~z70wang/publications/uqi.html

It would be nice to have some benchmarks to compare with some other implementations.

http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.73.4597&rep=rep1&type=pdf

Coveralls seems to be a nice tool for supervising code coverage. It would be nice to have it running together with badge displayed in README.

Visual Information Fidelity (VIF) is common choice for full reference image quality assessment. VIF has shown to better reflect analysis performed by humans.
https://en.wikipedia.org/wiki/Visual_Information_Fidelity

Link

FSIM: A Feature Similarity Index for Image Quality Assessment