Analysis and comparison of similarity measures and indices for image quality assessment

SILVIA MARÍA OJEDA; MARÍA LUCÍA PAPPATERRA

Guayaquil

Congreso; International Conference on Robust Statistics (ICORS) and The Latin American Conference on Statistical Computing (LACSC); 2019

Escuela Superior Politécnica del Litoral (ESPOL)

The amount of digital imagery is rapidly increasing every year to the extent that subjective assessment of image quality has become virtually impossible due to time and cost constraints. Furthermore, in digital image processing there is a need to compare the performance of different image processing algorithms by comparing the quality of their output images. To overcome these and similar problems many objective image quality indices were developed. Although many of these indices have been proposed, they stem from different theoretical frameworks, and thus their application scenarios are different, and they may serve different purposes. To the best of our knowledge, this is the first work to compare a large number of indexes, analyzing their differences in performance and assessing which index is more suitable in which application scenario. We select, analyze and compare several full-reference indices and measures: the mean square error (MSE) and root mean square error (RMSE), the signal to noise ratio (SNR), peak signal to noise ratio (PSNR) and weighted signal to noise ratio (WSNR), the noise quality measure (NQM)and visual information fidelity (VIF), the universal quality index (UQI), the structural similarity index (SSIM) and multi-scale structural similarity index (MSSIM), the gradient magnitude similarity mean (GMSM), the gradient magnitude similarity deviation (GMSD) and the codispersion coefficient based CQ-Index. Our Python implementation of all these indices can be found at https://github.com/lucia15/IQA-metrics. We use Kendall?s Tau and Spearman?s Rank Correlation Coefficient and other nonparametric correlation tests and methods in order to determine the best procedures for comparing digital images, for their mathematical and statistical properties and their ability to emulate the Human Visual System (HVS).