Research On Intelligent Editing Algorithm Of Highlight Video For E-sports "Game For Peace"

In recent years,the field of video summarization has developed rapidly,but there is little research on highlight editing of gaming e-sports videos.Manual editing of videos cannot effectively cope with the tens of thousands of new videos every day;Game background logs cannot be obtained by ordinary users,the end-to-end game video highlight recognition model based on deep learning is not yet ideal,lacks controllability,and is difficult to implement in actual scenes.This paper takes "Game For Peace" as an example,and proposes a set of methods to automatically locate and identify the key object content information of the video,and complete the automatic semantic annotation and highlight editing of the game video,based on deep learning object detection framework YOLOv3 and a series of targeted traditional morphological methods,according to the high accuracy of deep learning and the fast speed of traditional image processing without training.The main contents of this paper are as follows:(1)In order to solve the problems of slow manual editing of the game video and difficult access to the backstage log of the game for "Game For Peace",propose a process for automatic editing of game videos.The process includes four steps:video acquisition and preprocessing,video information extraction and recognition,video frame processing and fusion,and uploading to database.The scheme of video information extraction and recognition makes full use of the advantages of traditional image processing algorithm quickly and without training and deep learning methods to locate high positioning accuracy,and can efficiently and accurately complete the task of semantic recognition of game videos.(2)Given that the game scene digital recognition rate is low,the speed is very slow.This paper proposes a digital recognition method based on YOLOv3,which regards digital recognition as a detection problem,combining pruning optimization YOLOv3 and heuristic merging strategy to complete the digital recognition task.Experiments show that this method can increase the recognition accuracy to more than 99%,and the running speed on the cpu is maintained at about 50ms per image.(3)In the extraction and recognition of video information,take "single residual blood anti-kill highlights" editing as an example,enumerate specific sub-module tasks including blood bar calculation,teammate status monitoring,and text broadcast,and give Specific scenario solutions.Aiming at the problem that traditional method locates inaccurately,a unified "Game For Peace" interested object detection framework YOLO-onehead is proposed,and introduce techniques such as deep PAN to improve model accuracy.Finally,a bottom-up game video highlight editing pyramid paradigm and three-tier architecture are summarized,which can provide ideas for the rapid deployment of other games.