Distortion Prediction for Video Quality Optimization over Packet Switched Networks

Scheduling techniques are often deployed at the network edge to maximize the quality of the video communication while satisfying a given constraint on the maximum high priority usable bandwidth. For the case of video communications, the importance of each packet, in terms of the distortion that would be caused by its loss, can be used to decide which packets should be prioritized in order to maximize the expected video quality. However, the resulting performance strictly depends on the length of the video stream segment considered, at each time instant, by the distortion-aware scheduling algorithms. This work focuses on improving the performance of those algorithms by predicting the characteristics of the near-future part of video streams, exploiting the short and long term dependencies in packet distortion values. This approach could be particularly valuable in live video scenarios where accessing the future part of the video would imply the insertion of an additional delay. Several distortion prediction models are derived and their performance evaluated with actual scheduling algorithms. Results show that the best predictions are provided by neural network models, yielding significant improvements of the video communication quality.