Research On Short Video Content Anaylysis Algorithm Based On Deep Learning

With the rapid development of the Internet and the rapid improvement of smart phone technology,a large number of short videos are shared via social platforms every minute.Therefore,video content analysis is a very important and popular work for machine learning and artificial intelligence.For a large-scale user-originated production of video content,how to conduct a full-scale analysis is a very difficult matter,how to filter out bad illegal content from a large number of user-issued short video,select high-quality video for sharing with other users,improve the entire user The quality of distribution platform video is also a top priority.This article analyzes the visual aspect and the auditory aspect of the video from the perspective of the video,audio,and publisher avatars from the user's visual and auditory perspectives.It uses multiple deep learning models to model these three domains and establishes multiple sets of depths.The learning model performs a comprehensive analysis of the video.First,based on the OPEN_NSFW network structure modeling,improvement,to determine whether the publisher avatar involved pornography,improve the accuracy of the discrimination,reduce the rate of misuse.Then,the video content is analyzed.Firstly,1)the inception-v3 model structure is used to extract the visual features of each frame;2)Fully-connected network structure is established respectively,SVM model is supported by vector machine,and Long-short-time recursive neural network(LSTM).This feature sequence is learned and expressed as a corresponding video category.In order to achieve higher accuracy and accuracy,one-layer LSTM,multi-layer LSTM,and LSTM with Dropout layer have been established on the basis of the LSTM structure of the neural network of the length and duration.Contrasting and analyzing the experiments to select the optimal model.3)Further adopt feature selection,add feature dimension information for each video,and better classify the video.Finally,analyze the audio in the video,use FFmeg to extract the feature vectors of the audio,adopt the idea of migrating learning,establish the classification model,and classify the audio into three categories: music,dialogue,noise,and adjustment of parameters.Accuracy.Then add the audio classification results to the video base dimension information to help the video better identify and classify.The three models that have been established have all been online,improving the user experience of hundreds of millions of users of the Weibo video recommendation service,and selecting higher quality videos to be sent to microblog short video users.