Research On Video Classification Method Based On Deep Learning

In order to effectively manage video data and extract important information in the video,automatic video classification technology has become the main way to solve this problem.Video data is composed of image frames,and the relationship structure of its internal information is relatively complex.Based on the characteristics of traditional manual design,it cannot effectively represent the complex information in the video data.Using deep learning technology can extract more complete feature information and improve the accuracy of video classification.Therefore,this study has studied video classification methods based on deep learning.The main work is as follows:(1)Aiming at the problem that the fixed-length video sequence cannot completely cover all the motion information of the video,this paper proposes a video classification method based on feature fusion of three-dimensional convolutional neural network.Based on the three-dimensional convolutional network model,this method extracts video frame sequence features with different time scales,and perforns weighted fusion on the features of the fully connected layer.Comparing different fusion methods on the data set UCF101,the experimental results show that the video classification accuracy of the back-end weighted fusion method is high;then the parameters of the weighted fusion are determined through experiments,thereby constructing a video classification network of feature fusion of different scales model.Experimental results show that this method is more effective than mainstream methods for video classification.(2)Aiming at the impact of video semantic changes on video classification results and how to improve intra-class similarity and inter-class dispersion in video classification,a multi-channel convolutional network video classification method based on deep metric learning is proposed.This method designs a multi-channel convolutional video classification network based on the network model of different scale features fusion.In order to enable the network to learn the intra-class similarity and inter-class dispersion,the interval assignment function based on the negative sample to the semantic distance is proposed in the metric learning structure,so that the network pays more attention to difficult samples.Simultaneously perform metric learning and classification tasks during training.Experimental results show that this method can improve the accuracy of video classification.(3)In order to further improve the accuracy of the classification results,a video classification method based on multiple convolutional networks and LSTM is proposed.This method extracts the spatial features of the video image based on the model of the multi-channel convolutional network;then uses the LSTM model to further obtain the features with temporal dynamic information,thereby obtaining more video timing features;finally,the Softmax classifier outputs the classification Results and classification accuracy.Experimental results show that this method can further improve the classification accuracy.