| Due to the rapid development of e-Sports video live broadcast in recent years,the detection of e-Sports video highlight has received more and more attention.E-sports video highlight detection is designed to retrieve a small part of the highlights of the team battle between the both sides in the game.However,the current video highlight detection for e-Sports depends on a large extent on human resources,that is,manual editing and uploading,which must consume huge manpower and financial resources.Therefore,it is a very popular research topic to use the relevance of video content in the time and space dimensions to automatically retrieve highlights and build a highlight detection model for e-Sports videos.There are very few deep learning algorithms applied to the problem of highlight detection in e-Sports videos,and other information,such as audio information,is required in addition to video information.In addition,existing algorithms in e-Sports video highlight detection often use heavyweight feature extraction algorithms or complex LSTM networks to extract contextual features in the time dimension.These algorithms are difficult to apply to the realtime e-Sports video highlight detection system.In order to solve the above problems,this thesis has carried out the following main research contents:(1)The thsis only relies on the video information to carry out the research on video highlight detection task of e-Sports,and designs an end-to-end video highlight detection model DENAN for e-Sports based on the attention mechanism.According to the structure of the eSports video,this model can obtain the strongly correlated spatio-temporal long-distance dependency,and it performs well in the task of highlight detectionof the e-Sports videos.A large number of experimental results on the NALCS and LMS League of Legends datasets can also prove the effectiveness of the DENAN model.(2)In order to prove the influence of the loss function on the detection results of DENAN,this thesis focuses on comparing the effect of the model trained only with the cross-entropy loss function and trained with both the cross-entropy loss and the triplet loss.Experiments show that the detection effect of combining the cross-entropy loss and the triplet loss has been significantly improved,indicating that the triplet loss function can effectively improve the detection performance of the model.In addition,With the help of triple loss function,the DENAN model proposed in this thesis has strong discriminative power and robustness.(3)On the basis of the DENAN model proposed,this thesis further compresses the model,and proposes a lightweight video highlight detection model QDENAN for e-Sports based on the attention mechanism.In this thesis,we discuss and experimentally verify the video highlight detection model for e-Sports based on the binarization,ternary or multi-value quantization network.In the end,this thesis found that the QDENAN model based on 8bit Do Re Fa-Net multi-value quantization performed well.Although the F1-Score of the DENAN model based on 8bit Do Re Fa-Net multi-value quantization is 72.59%,which is3.67% lower than the F1-Score value of the full-precision floating-point DENAN model of76.26%,but the model size is compressed from 12.95 M to 3.88 M,proved that the model compression is of great significance for improving the hardware adaptability of the model.(4)Based on the previous research work,this thesis adopts the Koa+My SQL+Bootstrap technical solution to construct a lightweight e-Sports video highlight detection system.Through this e-Sports video highlight detection system,those who work on e-Sports video editing are rescued from the boring and long work.The system will provide users with streamlined and wonderful e-Sports video viewing services in two ways: video mode and report mode,saving gaming enthusiasts’ time to watch e-Sports videos and meeting their daily needs. |
PDF Full Text Request