Avaliação de qualidade de vídeo utilizando modelo de atenção visual baseado em saliência

Abstract

Video quality assessment plays a key role in the video processing and communications applications. An ideal video quality metric shall ensure high correlation between the video distortion prediction and the perception of the Human Visual System. This work proposes the use of visual attention models with bottom-up approach based on saliencies for video qualitty assessment. Three objective metrics are proposed. The first method is a full reference metric based on the structural similarity. The second is a no reference metric based on a sigmoidal model with least squares solution using the Levenberg-Marquardt algorithm and extraction of spatial and temporal features. And, the third is analagous to the last one, but uses the characteristic Blockiness for detecting blocking distortions in the video. The bottom-up approach is used to obtain the salient maps, which are extracted using a multiscale background model based on motion detection. The experimental results show an increase of efficiency in the quality prediction of the proposed metrics using salient model in comparission to the same metrics not using these model, highlighting the no reference proposed metrics that had better results than metrics with reference to some categories of videos.