Research On Ad Insertion Algorithm Based On Multiple Video Semantics

The rapid development of the Internet and the rapid increase of global video netizens have led to the rapid development of video platforms and injected great momentum into the development of the video economy.Advertising revenue is the source of video economic benefits and can bring huge commercial benefits.However,uncontrolled advertising will seriously affect the user’s viewing experience,which not only damages the interests of users,but is also detrimental to the development of video platforms.Studies have shown that when there is a connection between the advertisement content and the video content where the advertisement is inserted,it can improve the user’s acceptance of the advertisement and even arouse the user’s interest in the advertised product.So the goal of this paper is to build an ad insertion system based on video semantics.Video semantics are rich and diverse,and this paper bases ad insertion on both object semantics and action semantics information.Firstly,the object semantics contained in the video is used as the entry point for ad screening.The one-stage object detection algorithm is used to extract the video object semantics.The algorithm network is divided into three parts: feature extraction,feature fusion and recognition detection.Feature extraction uses convolutional neural networks to extract image features.Feature fusion fuses feature maps at different scales in the feature extraction network to unify feature information at different network depths.Recognition detection uses convolution and full connectivity to perform boundary regression on the prediction frame of the target while classifying the items,using one network to obtain both class information and location information of the target.Secondly,the main content of the video is understood from the action semantics for relevant ad insertion.In this paper,we study a multi-stream based action recognition algorithm.The characteristics of three data streams,RGB,optical stream and human skeleton,are combined to improve the recognition rate of behavioral actions.The information of the three data streams complement each other,and the recognition accuracy of the network is improved by fusing the three data streams for judgment.Optical streams describe motion features using inter-frame difference information.The human skeleton was data enhanced on the basis of the original image,highlighting the location of the human body in the picture and the human posture,and downplaying the effect of human appearance.The information of the three data streams complement each other,and the recognition accuracy of the network is improved by fusing the three data streams for judgment.Finally,this paper designs the advertisement insertion system based on the object semantics and action semantics of the video.This paper designs an ad insertion system based on video content,which draws closer the connection between video content and ad content,and presents ads in a form similar to pop-up windows to arouse users’ interest while giving them a minimal sense of viewing disgust and achieving optimization of ad insertion.