改进GAN网络在生成短视频的应用研究  

Research on Application of Improved GAN Network in Generating Short Video

作　　者：于晓明[1] 黄铧 YU Xiao-ming;HUANG Hua(College of Electronic Information and Artificial Intelligence,Shaanxi University of Science and Technology,Xi’an 710021,China)

机构地区：[1]陕西科技大学电子信息与人工智能学院,西安710021

出　　处：《计算机科学》2021年第S02期625-629,共5页Computer Science

基　　金：陕西省科技厅项目(2014KRM80);咸阳市科技局项目(2013K15-07)。

摘　　要：在研究生成对抗网络(GAN)生成动态图像时,经常出现前后帧图像内容中的部分物体颜色不一致和生成的细节不自然等问题。针对当前生成视频的不理想问题,采用的主要方案是分别对GAN网络中的生成器和判别器进行改进,具体表现在两个方面:一方面是在生成器中对视频的前景和背景分别建模,并且使用多重空间自适应归一化(Multi Spatially-Adaptive Normalization,M-SPADE)算法;另一方面是在判别器的选取上使用双视频判别器(DVD-GAN),然后在Kinetics-600数据集进行训练,实验后的结果分别比对F-Vid2Vid,WC-Vid2Vid等生成方法。实验结果证明了对GAN网络改进的方法在处理生成短视频的前后帧颜色不一致的问题和细节上有着不错的效果,生成的图像相对的更加清晰。In the study of the dynamic image generated by GAN,there are many problems,such as inconsistent colors of some objects and unnatural details of generated images.In order to solve the problem of unsatisfactory video generation,the main schemes adopted are to improve the generator and discriminator of GAN network respectively,which are shown in two aspects.On the one hand,the foreground and background of the videos are modeled separately in the generator and the Multi Spatial-Adaptive Normalization(M-Spade)algorithm is used.The other aspect is the use of dual video discriminator(DVD-GAN)on discriminator selection,which trained on Kinetics 600 dataset.The experimental results are compared with F-VID2VID,WC-VID2VID and other generation methods.The results show that the method of combining the two methods has a great effect on the problem of color inconsistency before and after the short video and the details processing,and the generated images are relatively clearer.

关 键 词：生成对抗网络 合成短视频 前景-背景图像建模 多重空间自适应归一化 双视频判别器 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]