Grebenyuk V.V. Methods of computer assessment of multimedia data quality based on non-reference methods - Qualifying scientific work on the rights of the manuscript.
The dissertation on competition of a scientific degree of the doctor of philosophy on a specialty 123 Computer engineering. - State University of Telecommunications of the Ministry of Education and Science of Ukraine. - Kyiv, 2021.The thesis is devoted to research on digital processing and multimedia. Much attention is paid to improving the quality of, for example, videos. Evaluating the quality of images and videos is an integral part of the processes in filtering and restoring video, cleaning them from defects such as Gaussian noise.
For research in this work, such sections of digital image processing as image segmentation, border detection, color space transformation, morphological image processing are essential.
The current state and prospects of development of digital processing of multimedia data are analyzed, the bases of the methods providing an estimation of quality of images and video materials are defined.
It was found that when creating, processing, transmitting and playing multimedia, in the vast majority of cases it is impossible to obtain original material to compare and detect distortions at each stage of the content life cycle, so image quality assessments were performed in the absence of a sample.
The analysis and comparison of not only non-reference methods, but also methods that require a reference copy of the data to compare and detect distortion in images or videos.
In the dissertation the actual scientific problem of development of methods and computer estimation of quality of multimedia data on the basis of non - reference methods is solved.
The method of searching and evaluating compression artifacts on the images, which differs from existing treatments using search blocking during combination of perceptual and pixel measures that can better be assessed compression artifacts in the images. This solution allowed to increase the reliability of the obtained results by 1-2% and to ensure the clarity of the range of initial metrics.
A method of non-reference assessment of image quality has been developed, which differs from the existing ones in that the metrics are calculated by statistical analysis of pixel brightness by individual color channels based on the calculation of the coefficient of variation and size of rectilinear areas of equal intensity by geometric halftone image. Such approaches not only allowed to evaluate and compare images for the presence of compression artifacts, but also allowed to use an algorithm to analyze and evaluate different types of noise in images.
For the first time the method of detection doubling artifacts in images and video frames, which is based on the use of methods detector boundaries Canny and Prewitt, allowing to identify and estimate the doubling of the image object and the magnitude and direction of blurring obtained by filming fastmoving object with a high shutter speed. By including the algorithm binarization images allowed to realize autom second choices IR sensitivity threshold, which in turn can increase the accuracy of the artifacts doubling to 10 % compared with the metric without automatic selection of the sensitivity threshold.
The method of calculating the overall assessment of the quality of video sequences, based on an analysis of three variables: the average of estimates of video frames, average of 10% of the worst grades and average of 1% worst estimates, which allows using the same metrics to evaluate quality of the video, not individual its parts.This approach, in contrast to the existing ones, allows to understand how many staff have high quality scores, and by giving preference to low scores allows you to estimate the number and value of ratings for low quality staff.
To confirm the reliability of the obtained scientific results with the help of software packages in C # and MATLAB languages, image and video files were processed. The developed technique of calculation of the general estimation of quality of video sequence allows to increase by 1-2% an estimation of quality of video, including as at a small level of distortions. Also, tests have shown that the algorithm with automatic selection of the threshold of sensitivity of filtering based on binarization of images is 10% more accurate in estimating artifacts.
