Photorealistic Synthesis Of Human Animation Based On A Single Image

Since it is more convenient to capture a single image of a human body than an image sequence,a single image-based photo-realistic synthesis of human animation has a wider range of application scenarios and has attracted much attention.However,because a single image lacks three-dimensional information and complete human body texture information,the synthesis of realistic human body animation based on a single image is extremely challenging.The existing algorithms have the shortcoming of high complexity,limited adaptation range,and low photorealistic.This thesis proposed a human body animation synthesis algorithm which is based on the reconstruction and rendering of the human body 3D model.From the view of the photo-realistic of resultant animations and the algorithm’s efficiency,the proposed algorithm firstly constructs a parameterized model of human body which is equiped with complete textures and more detailed geometry,and then 3D motion is applied and drives the model to move to generate a highly realistic human animation.The major tasks include:First of all,a 3D human body model reconstruction method based on SMPL-X parametric model is proposed.Based on the SMPL-X parameter model of the human body contour information in the input image,the three-dimensional geometric models of the front and back sides which are consistent with the object contour are generated.Then,a mesh splicing fusion algorithm based on B-spline interpolation is used to splice the front and back three-dimensional geometry.Finally,in order to restore the correct hand geometry,the wrong hand geometry on the reconstructed model is replaced by the correct hand geometry on the standard SMPL-X parametric model by the mesh stitching fusion algorithm based on B-spline interpolation.Secondly,back-face texture generation algorithm based on the Conditional Generative Adversarial Nets is proposed.The second important part of the framework is to solve the problem of the lack of human back texture in a single image.Since the Obverse and reverse of the human body have the same contours and many of the same visual features from the same perspective,in this paper,based on the Conditional Generative Adversarial Nets,a back-face texture generation network is trained to act as a transform function to generate human back-face texture image from the input front image.Finally,based on the input motion data,the pose parameters of the reconstructed human model with complete texture details and geometric details are changed to control the model motion and render the animation.This thesis has conducted sufficient experiments on existing public data sets and data collected on the Internet.The experimental results show that the human geometry with complete texture generated by this method can be driven by 3D motion data to generate a highly realistic 3D human animation.In summary,the main contribution of this paper is to propose a parameterized 3D human body reconstruction algorithm based on a single image,design and train a backside texture generation network based on Conditional Generative Adversarial Nets.And based on the above algorithm,a framework of a realistic human animation synthesis algorithm based on a single image is constructed.