Research On Adaptive Bitrate Streaming Algorithm Based On Machine Learning

With the continuous development of the domestic digital wave,new applications of mobile Internet are emerging one after another.Especially those that rely on streaming media technology,such as Douyin,Kuaishou and various live broadcast platforms,have created explosive growth of people’s demand and profoundly changed the living styles of the Chinese people.Correspondingly,the widespread application of network video has also promoted the development of Internet technology,and has continuously put forward higher requirements for the development of streaming media technology.In the field of network video,streaming media transmission technology has always been a research hotspot.The dynamic adaptive streaming media technology(DASH)based on HTTP is the streaming media transmission technology with the highest market share.Compared with traditional video-on-demand systems,DASH can be based on the user’s network conditions to help users choose the appropriate bit rate in real time,improve overall fluency and clarity,and avoid jams.Thus,users can achieve ideal playback effects under different network environments.In the early days of DASH development,because there was no unified standard,many manufacturers proposed their own solutions for their own business needs,resulting in numerous corresponding agreements and poor universality.The MPEG-DASH standard was born under this circumstance,and has become one of the mainstream protocols of streaming media transmission technology.Based on the research foundation of the MPEG-DASH protocol,this paper has carefully studied and reproduced three classic dynamics,i.e.,bandwidth-based rate-adaptive algorithm,buffer-based rate-adaptive algorithm,and hybrid model-based rate-adaptive algorithm.Although the principles of these algorithms are simple,easy to understand,easy to use,and effective,they also have obvious shortcomings.For example,a bandwidth-based rate-adaptive algorithm will lose its effect under a highly volatile network environment,and a buffer-based rate-adaptive algorithm cannot show its advantages when the bandwidth is relatively stable.Based In this situation,this article attempts to find an algorithm that can achieve an ideal performance under different network environments.In view of the rapid development of deep learning networks in recent years,which has obtained great success in many fields such as speech,image and natural language processing,this article attempts to apply deep learning to dynamic bit rate adaptive scenarios,and proposes a bit rate adaptive algorithm of the MPEG-DASH protocol based on LSTM(Long ShortTerm Memory).Due to the limitation of experimental computation resources,this paper selects a suitable network data set to train the deep learning model,and obtains an LSTM-based rate-adaptive algorithm that can adapt to multiple network environments.This article constructs a complete DASH experimental environment with a server and a client and based on the MPEG-DASH protocol on two computers,and finally implements the three existing models mentioned in the paper as well as the trained deep learning model.Specific playback data in the form of printing logs are obtained.The analysis of log files shows that,compared with the existing three classic rate adaptation algorithms,the algorithm proposed in this paper has stronger advantages in fluency,smoothness and stability,and can be well used in different network environments.In both cases,satisfactory results can be achieved.