Overall quality prediction for HTTP adaptive streaming using LSTM network

Abstract

HTTP Adaptive Streaming (HAS) has become a widely adopted method for delivering multimedia content. However, the variability in network bandwidth often results in fluctuating video quality during streaming, posing a significant challenge in evaluating the overall quality of a streaming session. This study proposes a machine learning-based approach to predict the overall quality of a streaming session by analyzing each segment's features.

Two feature sets are explored for this purpose. The first feature set includes segment quality, content characteristics, stalling duration, and padding, while the second set comprises bitstream-level parameters, stalling duration, and padding. These features are input into a Long Short Term Memory (LSTM) network, which captures temporal dependencies between quality impairments and stalling events. A linear regression module then uses the LSTM network's outputs to predict the overall quality.

Experimental results demonstrate that the proposed approach achieves high prediction accuracy and surpasses seven existing methods. Notably, the second feature set proves to be both efficient and effective. The source code for the proposed method is publicly available.