Video Content Recognition Technology Research Based On Deep Learning

With the development and popularization of the Internet and 4G network technology,and dominant social network give more and more attention and promotion to video business,such as Facebook,Instagram and Snapchat,network video business is booming on the way.However,a growing number of video and user groups,as well as the rich variety of video content take a great challenge to the supervision and management of video content.Thanks to the breakthrough progress in deep learning in intelligent recognition,video content recognition technology based on deep learning has gradually become the main technology of video content recognition and analysis.Therefore,on the basis of deep learning,this thesis introduces the attention mechanism and makes full use of the temporal characteristics of video to study the accurate and efficient video content recognition technology.The core of the video content recognition is features extraction,and deep learning has powerful ability to extract feature.In order to further research the video recognition based on deep learning.The major works of this thesis are listed as follows:First of all,combined with the human visual perception research and the advantage of LRCN in video content recognition,this thesis proposes a LRCN model based on attention mechanism from the global to consider the video content.The proposed model simulates the attention characteristics of the human brain in the deep learning model,makes the attention of the model on the effective area of the whole video,and eliminates irrelevant information on video content recognition.The proposed model is based on the selective attention weight,which is used to assign a larger weight to the region related to the video theme,and the irrelevant area is assigned a smaller weight to extract the discriminative temporal features.Then,to make full use of the time characteristic of the video,this thesis proposes a LRCN model based on BLSTM.The proposed model adopts BLSTM network to capture the context information of the video,and extracts abundant temporal features.BLSTM network is used to extract the forward and reverse temporal features for video content recognition at the same time,which makes full use of the time characteristics of the video content information.Finally,the two proposed model are simulated and verified by the deep learning open source framework Tensorflow,and the experimental data adopts the video content recognition dataset HMDB-51 and UCF-101.Experimental results show that the LRCN model based on attention mechanism and the LRCN model based on BLSTM network can both improve the accuracy of the video content recognition effectively,and the LRCN model based on BLSTM network can converge fast in training,which improves the efficiency of the model.At the same time,this thesis visualizes the attention weights of the LRCN model based on attention mechanism,and analyses the effect of attention weights on the video content recognition.