Métodos sem referência baseados em características espaço-temporais para avaliação objetiva de qualidade de vídeo digital

Abstract

The development of no-reference video quality assessment methods is an incipient topic in the literature and it is challenging in the sense that the results obtained by the proposed method should provide the best possible correlation with the evaluations of the Human Visual System. This thesis presents three proposals for objective no-reference video quality evaluation based on spatio-temporal features. The first approach uses a sigmoidal analytical model with leastsquares solution using the Levenberg-Marquardt method. The second and third approaches use a Single-Hidden Layer Feedforward Neural Network with learning based on the Extreme Learning Machine algorithm. Furthermore, an extended version of Extreme Learning Machine algorithm was developed which looks for the best parameters of the artificial neural network iteratively, according to a simple termination criteria, whose goal is to increase the correlation between the objective and subjective scores. The experimental results using cross-validation techniques indicate that the proposed methods are correlated to the Human Visual System scores. Therefore, they are suitable for the monitoring of video quality in broadcasting systems and over IP networks, and can be implemented in devices such as set-top boxes, ultrabooks, tablets, smartphones and Wireless Display (WiDi) devices.