Design And Implementation Of Automatic Video Description Based On Deep Learning

With the rapid development of the Internet,online multimedia information has grown dramatically.Short videos and live video broadcasts have been particularly hot in recent years.The video contains a wealth of information.Describing the video in natural language is extremely important to understand the video and retrieve it on the Web.The video description refers to generating a corresponding sentence for the given video by observing the content contained in the video,including the objects in the video and the activities of the object.Faced with massive video,if the video is described in a manual way one by one,the cost is very high.Using computer technology to analyze the video features and combining with the natural language processing method is an effective method to solve the "semantic gap" problem.This thesis describes video data based on Deep Learning,a coder-decoding model based on deep learning is proposed.Time attention mechanism and space attention mechanism are added in this model.Characterization of the semantic representation of key information are extracted by convolution neural network including video of the activities,such as the object of activity.The neural network with C3 DNet can improve the activities and activity of video object recognition accuracy,using VGGNet network extraction of video data in the surface characteristics of the object,and then use the cycle neural network(LSTM length memory network)semantic representation to generate natural language description of extracted.Finally,we train and test the model with YouTubeClips and TACoS-MultiLevel video data set.Using BLEU,METEOR and CIDEr three indicators to assess the model.Then,Compared with the LSTM-YT,S2 VT,MM-VDN,TA,LSTM-E and CRF-T,CRF and-m LRCN,our method is superior to other methods,and meet the needs of the system can meet the demand.