Features Extraction for Live Streaming Video Classification with Deep and Convolutional Neural Networks

These days, the high-quality live video streaming platforms are very widely used by lot of people all over the world. The demand for such services has increased even more in current pandemic conditions, so it is important to maintain certain level of Quality of Service for live streaming video. This article aims at pre-processing live streaming video-traffic packets for Deep Learning application of classification of different quality mode (Resolution and framerate) based only on statistical properties of network traffic Internet Protocol (IP) packets flow. The traffic intensity is used as a main traffic type descriptor with some additional transforms. The article presents some of classification accuracy analysis depending on multiple factors, such as: type of traffic pre-processing, length of traffic window, traffic intensity sampling time. The results show that there is a set of such parameters that leads to the best classification accuracy. This way, it may be possible for content operators (platforms) to provide differentiated services for their users. According to testing results, for Deep Learning application the direct traffic intensity samples allows achieving good accuracy of 94% and higher without any additional transforms. This, of course, depends on the classifier model parameters, and article provides insight into them as well.