基于上下文块的隐空间图像动画技术

Hidden Space Image Animation Method Based on Context Block

图像动画方法通常是指将驱动视频的运动“复制”给源图像,使源图像获得与驱动视频相同的运动,让静止图像动起来。然而,以往的图像动画方法通常需要提取标签、光流、语义图等结构信息,如果源图像和驱动视频具有较大的外观变化,就会难以保留源图像的身份信息,生成伪影和错误的运动。基于此,提出了基于上下文块的自监督隐空间图像动画方法,通过隐空间中的线性导航使图像动画,由隐空间编码的线性位移来生成驱动视频的运动。在编码部分引入上下文模块能够更加准确地对源图像和驱动图像进行编码,生成准确真实的图像动画。实验分析表明,该模型在VoxCeleb2、Celeb-V2和Ted-talk谈话数据集上系统地和显著地优于其他方法。

The image animation method usually refers to"copying"the motion of the driving video to the source image,so that the source image can obtain the same motion as the driving video and make the still image move.However,previous image animation methods usually need to extract structural information such as labels,optical flows,semantic images,etc.If the source image and the drive video have large appearance changes,it will be difficult to retain the identity information of the source image,and generate artifacts and wrong motion.Based on this,this paper proposes a self-supervised hidden space image animation method based on context block.The image is animated by linear navigation in the hidden space,and the motion of driving video is generated by linear displacement encoded in the hidden space.The introduction of context module in the coding part can more accurately encode the source image and drive image,and generate accurate and real image animation.Experimental analysis shows that this model is systematically and significantly superior to the other methods in VoxCeleb2,Celeb-V2 and Ted-talk conversation data sets.