Sistemnyj Analiz i Prikladnaâ Informatika (May 2017)
COMPARATIVE ANALYSIS OF NO-REFERENCE QUALITY MEASURES FOR DIGITAL IMAGES
Abstract
This paper presents results of a comparative analysis of 34 measures published in the scientific literature and used for evaluation of the image quality without a reference image. In English literature, they are called no-reference (NR) measure or measures NR-type. The first article, the term no-reference, was published in 2000 and each year a growing number of publications on new measures NR-type. However, comparative studies of such measures is not practically conducted. Such measures are very important for a) just made photo quality evaluation, b) assessment of image enhancement transformations and selection of their parameters (such as contrast and brightness adjustments, tone-mapping, decolorization and others). Publicly available image quality databases used for study no-reference quality measures (TID2013, etc.), contain 4-5 variants of images distorted by predefined transformations with unknown parameters. We presented six types of experiments to analyze correlation of the computed numerical quality values with visual estimates of the test images quality. Four of the experiments are new: comparison of images after gamma-correction and contrast enhancement with different parameters, as well as analysis of the retouched images and photos taken with different focal length. It was shown experimentally that no one of the known no-reference quality assessment measure is universal, and the calculated value cannot be converted to a quality scale, excluding factors influencing the distortion of the image. Most of the studied measures calculates local estimates in small neighborhoods, and their arithmetic mean is the quality index of the image. If the image contains large areas of uniform brightness, the measures of this type can give incorrect quality assessment, which will not correlate with the visual assessments.
Keywords
