Research on Future Video Frame Prediction Method for Dynamic Convolutional Neural Advection Generative Adversarial Network
AN Li-zhi1，HE Ping1，ZHANG Wei2，SHI Yu-yang1，TIAN Yu1
1(School of Artificial Intelligence，Hebei University of Technology，Tianjin 300400，China)2(School of Electronic and Information Engineering，Hebei University of Technology，Tianjin 300400，China)
Abstract：Aiming at the problems such as poor prediction accuracy and object structure information loss in the current video frame prediction model，a dynamic convolution generation countermeasure network is proposed.In the generator network，the image features of the input video stream are initially extracted by using convolutional long short-term memory network.Then the motion features in the video stream are extracted by using the convolutional dynamic neural advection.Finally，a set of predictive video frames are output after combining the above two features.In the discriminator network，a 3D convolutional network is used to accept all video frames at once.In the experiment，Adam method is used to optimize the parameters of the model and KTH and BAIR Robot Pushing data set are used as training data sets.The experimental results show that the dynamic convolution generation confrontation network is better than the variational generation confrontation network in terms of long-term video frame prediction accuracy and object structure information retention，as well as the subjective perception of the human eye.Its structural similarity measurement The index increased by 14.5%，and the index of learning to perceive image block similarity increased by 7.69%，and the generated predictive video was smoother，which has higher practical value.
安利智,何平,张薇,石钰阳,田宇. 动态卷积生成对抗网络的视频帧预测方法研究[J]. 小型微型计算机系统, 2022, 43(2): 278-284.
AN Li-zhi,HE Ping,ZHANG Wei,SHI Yu-yang,TIAN Yu. Research on Future Video Frame Prediction Method for Dynamic Convolutional Neural Advection Generative Adversarial Network. Journal of Chinese Computer Systems, 2022, 43(2): 278-284.
2022, Vol. 43 
