Research Of Video Classification Method Based On Multiple Deep Feature Fusion

With the development of computer science and multimedia technology,videos can be easily obtained and stored.Particularly,internet makes video becoming a major way for communication.The number of videos on the internet is growing rapidly which results in the challenge to understand and manage videos.Traditional video classification methods have been unable to meet the growth speed of the videos.Video classification is an important branch in the field of computer vision,and has wide application prospects.There is more motion information in the videos than the static images and characters.The hand-crafted categories,which focus on global video representations,do not obtain the key motion information of video.Recently,motived by the promising results of deep networks on the image analysis tasks,most of the researches focus on deep feature extraction based on the deep learning and convolutional neural networks.Therefore,this thesis bases on the deep features which are generated from deep networks,and aims to make full use of the deep features to generate a high accuracy.For the motion features extracted from the 3D convolutional neural network,this thesis proposes a multi-time scale framework to classify videos.In order to increase the diversity of features,our method utilize the deep features of multiple time scales and of different layers.Then,a novel deep feature encoding method is proposed to generate video representation.For the deep features extracted from the 2D and 3D convolutional neural network,the attention-based feature fusion method is introduced to take advantage of the attention context generated from the hidden state of the Long-Short Term Memory.The attention model improves the ability of deep features to filter the redundant information.To fuse the multiple features,a two-layer neural network with a softmax in the end is introduced to generate attention weight distribution over the different features.In order to validate the effectiveness of the proposed methods,sufficient experiments are conducted on the popular UCF101 dataset.The results of experiment show that the video classification methods proposed in this thesis can effectively solve the problem of existing methods and achieve good performance.