Short Video Classification Based On Multi-Model Combination

The scale of images and short videos on the Internet is growing,and the fast and efficient short video automatic classification algorithm can help people find the video content of interest more easily.At present,the deep convolutional network model is an effective method to realize automatic classification of short video,but the convolutional network of single model has insufficient generalization ability.Aiming at the above problems,this paper studies the short video classification method based on multi-model fusion.The model fusion combines multiple single models through different combination strategies,and finally outputs the prediction results.The main work of this paper is as follows:1.Short video database is established.Construct an action video dataset containing six categories of 600 different pedestrians.Each of these categories contains 100 video samples,each of which is approximately 15 seconds long and has a frame rate of 25 fps.For each video(training and test video),200 frames are saved as training set and test set,and the classification performance of the image is used as the classification performance of the corresponding video:Train set has 80,000 frames,and Test set has 20,000 frames..2.Design and implement video classification based on 3D convolutional neural network.Use the 3D convolution kernel to extract video time and space information.The network consists of an input layer,four convolutional layers,four pooling layers,four batch regularizations(BN layers),two fully connected layers,and an output layer.The training process uses data enhancement,Dropout,adaptive learning rate and other techniques,and the trained model achieves a classification accuracy rate of 84%.3.A multi-model fusion short video classification algorithm based on 3D convolutional network is designed.1 Multi-model fusion algorithm based on Averaging.Average output of multiple model predictions to determine short video categories.The experimental results show that the method can achieve an average classification accuracy rate of 85.1%;2 multi-model fusion algorithm based on voting method(Voting).Using the relative majority voting method,the model classification results are voted,and the category with the highest number of votes is selected as the final prediction result.The experimental results show that the method can achieve 88.0%average classification accuracy rate;3 based on stacking multi-model fusion algorithm.Short pre-training models are used to extract short video features,compose these features into new data samples and train the SVM classifier,and finally determine the short video category through the SVM classifier.The experimental results show that the method can achieve an average classification accuracy of 89.4%.