Design And Implementation Of Video Generation Algorithms Based On Neural Networks

The generation model based on neural network aims to make the neural networks generate the samples which conform to the distribution of training data by learning the training samples.As an important generation model,Generative Adversarial Networks has attracted more and more attention in the application of video generation.However,many models based on Generative Adversarial Networks are faced with training difficulties and complex network design when they are used in video generation tasks,so it is necessary to improve the relevant methods.In this paper,the structure of Variational Autoencoder and the Generative Adversarial Networks for generating latent variables are designed.In addition,we study the gradual change of the image corresponding to the variable in the latent space of the Variational Autoencoder and the feasibility of combining the Variational Autoencoder with the Generative Adversarial Network for the video generation task.Finally,this paper designs an algorithm framework for the "text video" generation task and verifies the feasibility and flexibility of the algorithm framework on MNIST handwritten font mobile video dataset,and verifies the superiority of the proposed algorithm by comparing with other methods.The experimental results show that the video generation framework,which combines the Variational Autoencoder with the Generative Adversarial Network,can generate videos with consistent content and smooth action for the input text,decouple the information between the object and action in the training sample through training,and generate correct videos for the description text not included in the training sample.The research in this paper can be applied in the field of multimedia,such as intelligent program production,human animation generation and so on.