A comparative analysis of non­reference assessment methods

Abstract

A comparative analysis of functions to assess image quality in the absence of a sample: no-reference (NR) measure or NR-type methods. The availability of NR-methods is very important for assessing the quality of streaming video such as television, game streaming, online conferences, web-chatting, etc. (because on the side of the recipient of the video there is no standard for quality comparison) and assessing the results of transformations aimed at improving video, and choosing the parameters of these transformations (brightness change, semitone and others). The human visual system (HVS) is able to visually assessing video quality, but If required to visually assess the quality of dozens or hundreds of videos or ranking them by quality level it will be needed a huge amount of time. Six types of experiments were performed to analyze the correlation of calculated quantitative estimates with visual assessments of the quality of the tested video files. Three of them are fundamentally new: comparing video after gamma correction and changing the contrast with different parameters, as well as blurring, which may be the result of defocusing the camcorder. A hybrid method (or reduced-reference (RR) measure) and a full-reference (FR) measure or FR-type method were also added for comparison. It has been experimentally shown that none of the studied non-reference methods of image quality assessment is universal, and the calculated assessment cannot be converted into a quality scale without taking into account the factors influencing the distortion of image quality. Moreover, all NR-type methods could not cope with the experiment of changing the contrast, believing that the best result is the most contrasting image but the original. Instead, the reference methods showed an excellent result (except one, which showed partial ineffectiveness). Also, it has been shown performance comparison between methods. It is shown that most of the studied methods calculate local estimates for each frame, and their arithmetic mean value is an estimate of the quality of the entire video file. If the video is dominated by large areas of uniform evaluation, methods of this type may give incorrect quality evaluations that do not coincide with the visual evaluations.