TWITCH LIVE STREAMING QUALITY OF SERVICE BASED ON DEEP FEATURE

Ye, Wenqian

Permalink

https://hdl.handle.net/2142/125049 Copy

Description

Title

TWITCH LIVE STREAMING QUALITY OF SERVICE BASED ON DEEP FEATURE

Author(s)

Ye, Wenqian

Issue Date

2020-05-01

Keyword(s)

Live streaming, Quality assessment, Deep neural networks.

Abstract

Live Streaming is getting more and more popular in people’s daily life. Twitch, a subsidiary of Amazon, is one of the world’s leading live streaming companies in the United States and they are facing a problem that there are no reliable metrics to assess the video quality of the client-side. The traditional metrics of video processing (L2/PSNR, SSIM, FSIM) disagree with human judgments. Also, the network compression mechanism (H.264) will introduce many unexpected artifacts. We try to use deep neural networks to extract the deep features to simulate human perception so that this system will provide valuable data for their future product improvement. To tackle this problem, we built an open-source library using OpenCV, Blipper, to simulate all possible video artifacts like buffering, coloring bleeding, contouring, etc. This system is also used to build our dataset and survey website for deep net training(on-going). In the training part, we will dive into many deep net architectures such as CNN, RNN, and RCNN to figure out the best model we will use. In the future, we hope to extend our project to broader industries, such as FaceTime on Apple iPhones platform, Zoom online meeting, and so on. This project will help many video companies to direct their products’ disadvantages and provide possible solutions.

Type of Resource

text

Language

eng

Owning Collections