Research On Pose-driven Human Silhouette Image Synthesis

Human image synthesis is a hot research issue in the field of computer vision.Current research on human image synthesis mainly focus on ordinary images with consistent exposure and clear background and foreground objects.Silhouette photos are very different from ordinary images,which have a clear subject and background.In silhouette images,the background is bright due to proper exposure,while the foreground is underexposed and appears dark.The dark foreground of the silhouette image and the bright background form a strong contrast,and most of the details of the foreground are lost,which brings a challenge to its synthesis.In addition,the shooting of silhouette photos is limited by many factors such as skill,time,location,weather,etc.The pose in silhouette photos taken in suitable scenes may be not ideal.We are eager to have the opportunity to interactively adjust the poses of the character.Human image synthesis can be end-to-end or phased.The phased method can not only improve the generalization ability of the network but also reduce the training difficulty.To this end,this dissertation proposes a phased method for the synthesis of human silhouette images.This method mainly involves person detection,human region segmentation,interactive deformation and image inpainting.The main contributions of this dissertation are summarized as follows:(1)Human image synthesis requires careful analysis of images to obtain the position,shape,and pose of the character.This requires the human body region segmentation and human pose estimation.To this end,this dissertation proposes a human silhouette image analysis method.The method combines threshold segmentation and human body detection to obtain the human region and estimates the human pose based on deep learning.Experimental results show that this method can automatically obtain accurate human body regions and human skeletons.(2)After analyzing the human silhouette image,this dissertation changes the posture of the human body and synthesizes the image.The dissertation proposes a pose-driven human silhouette image synthesis method.This method changes the human skeleton in an interactive manner,thereby driving the human body deformation,and uses the image completion method based on deep learning to fill the holes formed by the human body deformation.The holes are adjacent to dark foreground silhouette pixels and are extremely irregular.These two unfavorable factors will jointly affect the completion effect.Therefore,in this dissertation,the background image that needs to be completed is processed,the foreground is color replaced,and the image is completed,and then the foreground dark color is restored,and the guide filter is used to soften the edge.Experimental results show that this method can synthesize realistic silhouette images.(3)In order to further expand the availability of pose deformation and synthesize the silhouette photos of real people in various poses,this dissertation presents a data-driven method of the three-dimensional human body parameterized model.This method obtains the three-dimensional human body parameterized model of the input image based on the deep network,and uses the human body shape parameters and pose parameters obtained from the reference image to drive the human body deformation,and then synthesize the human body silhouette image.