A Novel Algorithm for Video Frame Prediction Based on Convolutional Neural Network

The fundamental support of the deep learning algorithm model comes from a large number of labeled data sets, but manually labeling a large number of data sets is Cumbersome, and it is easy to produce huge errors, which has also become a bottleneck in the development of computer vision. In this paper, video prediction has become a secret medicine to solve the bottleneck of the development of computer vision technology. The essence of video prediction is to predict and generate video frame by frame. The traditional LSTM algorithm has a large amount of calculation due to the internal full connection. ConvLSTM greatly improves the prediction performance. However, conventional ConvLSTM cannot guarantee the clarity of the predicted output image. This paper presents a novel ConvLSTM prediction model. Long short-term memory network is applied in the field of video prediction. ConvLSTM no longer simply extracts temporal features, but also captures spatiotemporal features. Compared with convolutional neural network, the feedforward neural network operations of ‘input-to-state’ and ‘state-to-state’ in traditional LSTM are replaced by convolution operations, and the row and column information of samples can be extracted through convolution structure. In addition, ConvLSTM transmits information from the previous layer of neurons to the three gates at the same time before the information is transmitted to the three gates, so that the neurons can capture the information retained for more time. The FitConvLSTM model proposed in this paper has a score of 95% on the SSIM evaluation function, and has achieved good results.