| With the gradual promotion of 5G communication technology,short video has set off a trend in people’s life.Video creation around characters is one of the mainstreams in the field of short video.Human pose transfer is a new research direction.Given an original picture and a target pose,the characters in the original picture will show the target pose.The technology can provide novel creative means and style for characters video.At present,there are two problems in the dance video generated by the mainstream short video platform.One is that when the key points are blocked or overlapped,the detection results of human key points are inaccurate,resulting in the missing and broken limbs of the human body generated by the posture transfer task;The second is that the current algorithm is not good at preserving the stylized features of human facial details,which leads to the generation of facial blur in the human image and the inability to identify its identity information.These problems reduce the quality of video generation and user experience.Therefore,the following research contents are proposed in this paper.Firstly,for the inaccurate detection results of human key points,the human body generated by posture transfer task has the problem of limb missing and fracture.In this paper,a human skeleton construction algorithm based on graph convolution neural network is proposed.Through graph convolution and cascaded feature fusion module,the algorithm fuses the human body information with the image context information of high,middle and low dimensions with different emphasis in an innovative way.Through this way,the algorithm improve the accuracy of human body image limb connection.Secondly,this paper proposes a facial detail supplement algorithm based on confrontation generation network.By adding style consistency discriminator to the traditional generation network and adding style consistency loss and semantic continuity loss to the original loss function,a more realistic and consistent facial picture can be generated.The above facial image will be used as the detail supplement of the human body image obtained after the human pose transfer to optimize the image generation results.Finally,through the comparison of visual images and common performance indicators,it is proved that the above two algorithms have good performance in guiding human images synthesis.This paper then designs and implements a dance video generation system based on human pose transfer.After analyzing the needs of the system,divide the functional modules required by the system,design a visual UI interface and apply the above two algorithms,the system improves the results of video synthesis and user experience. |