Analysis And Implementation Of Scene Classification For Live Video On Spark

Webcast is a new culture for video after traditional television broadcasting,its development has three stages.The first stage is websites such as Youku which can accept video uploaded by personal.The second stage is web clients such as YY which support live show.And the third stage is mobile webcast which can live broadcast at anytime and anywhere.The last two stages which have just experienced less than three years occupy the main market.According to incomplete statistics,there are more than 200 live platforms.And the Founder Securities predicted that the live market size will reach 60 billion at 2020.Webcast will become the third traffic center for mobile internet following Weibo and Wechat.The quality of the video is the most important issues of webcast all the times because of the lack of high-end equipment and professional post-production.Live platforms including Baidu,Ali and so on provide optimization for the quality of videos on the basis of the scene where the videos were recorded.It means that it is not fit when the recorder changes the scene.Therefore,this project which is implemented on Spark is proposed to provide evidence for dynamic adjustment for optimization of live videos by classifying the scene of video streams real-time.This system which was implemented on Spark Streaming classifies videos real-time and parallel.It was established on Spark and it used the thought of message queue.Video processing includes the transfer from video stream to frames,gray processing of the frames,histogram equalization processing,HOG feature extraction and optical flow feature extraction.Meanwhile,this project built a convolutional neural network(CNN)model named Multi-stream AlexNet(MSAN)which is based on AlexNet to receive multi-stream input.We use MS AN to classify the scene in this project and group continuous frames according to the minimum unit of video streaming.We use statistic of classification for frames in a group to determine the final classification of the minimum unit to realize the real-time classification of live video.We have trained a MSAN model whose average accuracy is 98%.The scene classification has been deployed on cluster and we have finished the unit test and system test.