Resolving Facial Image Defects Through the Integration of Segmentation and Fusion Networks

Generating images with dynamic expressions has become a key component of the social media industry. To enhance the realism, it is crucial to remove the potential distortion in the face image. Unfortunately, videos produced by PIrenderer often exhibit facial blur due to inadequate segmentation between foreground and background. In this paper, GrabCut and MODNet are used to post-process the videos generated by PIrenderer and then fuse them. Our proposed method can guarantee to reduce the influence of background on the face when generating dynamic expression videos. These post-processing steps optimize dynamic facial expression rendering, mitigate the face distortion problem, and ultimately produce more realistic video output.