Combining audio and video metrics to assess audio-visual quality

In this work, we studied the use of combination models to integrate audio and video quality estimates to predict the overall audio-visual quality. More specifically, an overall quality prediction for an audio-visual signal is obtained by combining the outputs of individual audio and video quality metrics with either a linear, a Minkowski, or a power function. A total of 7 different video quality metrics are considered, from which 3 are Full-Reference and 4 are No-Reference. Similarly, a total of 4 audio quality metrics are tested, 2 of which are Full-Reference and 2 are No-Reference. In total, we tested 18 Full-Reference audio-visual combination metrics and 24 No-Reference audio-visual combination metrics. The performance of all combination metrics are tested on two different audio-visual databases. Therefore, besides analysing the performance of a set of individual audio and video quality metrics, we analyzed the performance of the models that combine these audio and video quality metrics. This work gives an important contribution to the area of audio-visual quality assessment, since previous works either tested combination models only on subjective quality scores or used linear models to combine the outputs of a limited number of audio and video quality metrics.