Video Event Classification Based On CNN-RNN Neural Network

For a long time machine learning has been concentrated in image field.The deep learning model represented by CNN has achieved good results in image field,but in the video field,there is the time information between frames except the image information in the frame.Because the old model can't handle the time information very well,video event classification has not been able to achieve a major breakthrough.In recent years,the RNN has achieved good results in speech processing,which proves its ability to process time information.This paper combine CNN and RNN to solve video event classification problem.The main contributions of this paper work are as follows:First,through actual test,it is proved that CNN can extract the frame features needed in this paper,the features extracted by CNN are adjusted to make them more suitable for the model in this paper.Through the actual test,the feature vectors extracted from the adjustment experiment after network adjustment have lower redundancy and better timing information.Then,this paper discusses the problems of traditional RNN structure,describe the internal structure of two improved structures.Through the actual test,the feature vectors extracted after network adjustment have lower redundancy and better temporal information,two layer convergent network structure has better performance.Finally,this paper combine CNN and RNN to verify the feasibility of the proposed scheme.The results show that compared with single use CNN or RNN,combining the two networks can greatly improve network performance.At the same time,the performance of two improved RNN structures,LSTM and GRU,was compared.In this program,GRU performance is better than LSTM.The scheme is tested on three authoritative video databases.The results showed that the performance of this scheme reached the top level in the industry,it is proved that this scheme has a great improvement in performance compared with traditional key frame feature classification method,and has good applicability and robustness.