基於深度學習改善自適應影像串流系統之分析與研究

本論文主軸在於探討與設計基於機器學習的自適應串流系統，串流之基礎框架採用DASH，此為現今最廣泛被採用的方法之一，透過觀察Buffer作為控制串流Bitrate的依據，此過程即是DASH自適應調整影像品質 (ABR) 的過程，其目的是使QoE最大化，也就是盡可能提升整體使用者體驗品質，然而本論文認為傳統的串流控制方法並不適用於現今高畫質影像的環境，理由是基於Buffer的方法在畫質的選擇判斷上過於保守，特別是在處理Bitrate變動複雜度更高的影片，常出現失真、延遲、Rebuffer等現象；對此本論文提出導入深度強化學習來實現ABR控制，透過強化學習能，客觀地輸入更多參數，如網路吞吐量、延遲時間、SSIM等，更使用卷積神經網路、Temporal Difference學習與Policy Gradient，以增加訓練效能，並使訓練完成的模型提供更精確的行為模式，以提升改善串流過程中產生的失真、延遲、Rebuffer等現象。在追求更高畫質的未來，採用本論文提出的強化學習結合DASH來實現ABR控制方法，能有效評估各式影片中Bitrate變化，輔以監控Buffer、Throughput，來達到預測串流環境之變化，更靈敏地採取最佳的資源調配方案，同時保證影片觀看者的QoE。

關鍵字

機器學習；強化學習； MPEG-DASH ；自適應串流； Quality of Experience ；影像串流

並列摘要

The main point of this paper is to discuss and design an adaptive streaming system via machine learning, the basic framework of streaming in this study adopts DASH, this is one of the most widely used methods today, by observing buffer as the basis for controlling bitrate, this method is the process of DASH adaptive bitrate adjustment of video quality (ABR), the purpose of this method is to maximize QoE, that is to improve the overall user experience quality as much as possible, in short, it is to improve the overall user experience quality as much as possible；However, my paper considers that the traditional streaming control method is not suitable for today's high-bitrate video environment, he reason is that the Buffer-based method is too conservative in the selection and judgment of image quality, especially when dealing with movies with more complex bitrate changes, distortion, delay, re-buffer and other phenomena often occur. For this problem, this paper proposes to import deep reinforcement learning to realize ABR control. Through reinforcement learning, more parameters can be input objectively, such as network throughput, delay time, SSIM, etc., I also use convolutional neural network, temporal Difference learning and policy gradient to increase training efficiency and make the trained model provide a more accurate behavior pattern to improve and improve the distortion, delay, re-buffer and other phenomena generated during the streaming process. In the pursuit of higher image quality in the future, the reinforcement learning proposed in this paper combined with DASH is used to realize the ABR control method, which can effectively evaluate the Bitrate changes in various videos, supplemented by monitoring Buffer and Throughput to predict the changes in the streaming environment. Be more responsive to optimal resource allocation while maintaining QoE for video viewers

並列關鍵字

Machine Learning ； Reinforcement Learning ； MPEG-DASH ； Adaptive Bitrate Streaming ； Quality of Experience ； Video Streaming

參考文獻

[1] White Paper: Cisco Visual Networking Index: Forecast and Trends 2017 – 2022 White Paper, Cisco, San Jose, CA, USA, 2020. [Online]. Available: https://twiki.cern.ch/twiki/pub/HEPIX/TechwatchNetwork/ HtwNetworkDocuments/white-paper-c11-741490.pdf

Google Scholar

[2] White Paper: Cisco Annual Internet Report 2018–2023 White Paper, Cisco, San Jose, CA, USA, 2020. [Online]. Available: https://www.cisco.com/c/en/us/solutions/collateral/executiveperspectives/annual-internet-report/white-paper-c11-741490.html

Google Scholar

[3] Sandvine Global Internet Phenomena Report, Sandvine, Waterloo, ON, Canada, 2021. [Online]. Available: https://www.sandvine.com/phenomena

Google Scholar

[4] I. Sodagar, "The MPEG-DASH Standard for Multimedia Streaming Over the Internet," in IEEE MultiMedia, vol. 18, no. 4, pp. 62-67, April 2011, doi: 10.1109/MMUL.2011.71.

Google Scholar

[5] Google Team, 2021. Recommended upload encoding settings, https://support.google.com/youtube/answer/1722171?hl=en

Google Scholar

主題瀏覽