End-to-End Video Quality Assessment with Deep Neural Networks

The explosiveness of media consumption patterns over the last years has brought a new landscape of high competition that forces the involved companies to worry about the quality of service of the end-user. However, current methods fail to measure the quality of experience in a manner close to that of the end-user. This work presents an artificial intelligence-based system able to assess the video quality in an end-to-end manner and in a similar way to what a user would do. For such purpose, a novel method for generating samples that reflects the quality of video signals over real network conditions has been implemented. Additionally, a hybrid neural network was developed. This hybrid neural network is comprised of convolutional and recurrent layers in charge of extracting spatial and temporal features respectively. Results show a high ability of the proposed system to generate reliable estimations. More precisely, the system can get precision values up to 80 compared to that of humans during the same task, which can go up 89. Moreover, it has been proved that such results can be achieved without the need for long training processes or large datasets.