This paper evaluates the use of deep learning prediction models for assessing the quality of experience of video applications using input data from a single measuring point at the smartphone. Three architectures for deep neural networks, single task fully connected (FC), multitask FC and convolutional, have been implemented. A set of four relevant metrics recommended by ITU-T models has been considered as outputs for the prediction models. The amount of data received by the device, a parameter which can be easily extracted from the majority of smartphones, is used as input to the machine learning algorithms. For the convolutional neural model, the amount of transmitted data is also used as input. The predictions models have been applied to evaluate YouTube applications on smartphones, but could be generalized to other services. The overall achieved performance on the test set is around 90%, a very promising result as the generation of new labelled datasets is quite feasible at low cost thanks to the fully automated test setup developed.
Paper
Full text
Application of Deep Learning Techniques to Video QoE Prediction in Smartphones
Semantic Scholar · Computer Science · 2019
Abstract
This paper evaluates the use of deep learning prediction models for assessing the quality of experience of video applications using input data from a single measuring point at the smartphone. Three architectures for deep neural networks, single task fully connected (FC), multitask FC and convolutional, have been implemented. A set of four relevant metrics recommended by ITU-T models has been considered as outputs for the prediction models. The amount of data received by the device, a parameter which can be easily extracted from the majority of smartphones, is used as input to the machine learning algorithms. For the convolutional neural model, the amount of transmitted data is also used as input. The predictions models have been applied to evaluate YouTube applications on smartphones, but could be generalized to other services. The overall achieved performance on the test set is around 90%, a very promising result as the generation of new labelled datasets is quite feasible at low cost thanks to the fully automated test setup developed.