Research And Application Of Denoising And Summary Algorithms For First-Person UGC Video

With the rapid growth of the Internet and technological innovations in integrated media,as well as the proliferation of mobile devices,video media has become a mainstream medium and it has become increasingly easy for users to shoot videos.The concept of User Generated Content(UGC)began to emerge,and it became popular for the public to share their lives by shooting a variety of videos and uploading them to major video content platforms,including many human-driven,self-referential and documentary first-person videos.Due to the nature of user production and online platform delivery this type of video often has a large amount of complex types of noise and contains a lot of repetitive and redundant footage.The aim of this research is to apply a deep learning approach to the denoising and summary content extraction of first-person UGC videos.In the design of the denoising framework,a deep video prior-based denoising method is used in order to get rid of the dependence of deep learning on large-scale training data.In the design of the denoising framework,the video time-domain stability is improved and the denoising effect is enhanced by using multiple frames as the input to the network and by using optical flow networks to calculate the loss function of multi-frame fusion.In the design of the summary method,the original video is pre-processed using the denoising framework to improve,and the optical flow computation network is used to introduce motion information between adjacent frames of the video so as to effectively fuse optical flow features and image features.The module in key frame importance data regression combines Bi-LSTM and self-attention mechanism for video frame importance estimation,enabling the video summary model to more accurately measure the importance of video content.In addition,this study analyses the needs of self-publishing users and designs a diary app.users can use mobile devices such as mobile phones to shoot long videos and combine the proposed denoising and summarising algorithms to summarise the filmed videos.