A Study Of Pose Transfer Image Generation Technology

Pose transfer image generation technology aims to convert the given pose of the source person into the set target pose,simultaneously keeping the appearance and clothing texture of the generated character image as consistent as possible with the source character.The key issue in the study of posture transfer is the representation of human posture.There are different methods on the representation methods of a person’s pose.At present,the widely used two-dimensional pose representation based on human key points has the advantages of convenient acquisition and mature algorithms,which can directly calculate the accurate pose key points through the source person image.Other methods,such as person pose based on 3D representation,are difficult to obtain,which limits their application scene.However,there are some defects in the generation model using two-dimensional human pose key points.For example,due to the problems of sparse pose key points and less information,the effect of converting person images with large pose changes is poor,and in the process of image conversion task,more sparse pose representation can not ensure the alignment between the source image and the generated image.Preserving the appearance and texture of the generated person image is difficult.Therefore,based on the person pose transfer,this paper improves the constraints of the model and improves the consistency of the generated person image pose and appearance texture with the source person image in the case of the supervised model and unsupervised model respectively.The contributions are listed as follows:(1)Supervised Person Image Synthesis with Postural Guiding Attribute-Decomposed GAN is proposed.The network is improved based on ADGAN.In ADGAN,the lack of correlation of the coding of each component of the decomposition component encoder leads to the deviation of character texture generation,and the network is difficult to converge.Therefore,the pose information is added to the input of the decomposition component encoder,and the global pose information is associated with each decomposition component during computing,thus the network can learn the association relationship between each component.For the pose and texture of the generated person,the cyclic consistency loss is introduced to constrain respectively,to further improve the accuracy of pose and texture.(2)Unsupervised Person pose transfer algorithm based on image hidden space exchange is proposed.The network is improved based on Swapping GAN.In the original network,because the structure of ordinary images is difficult to express directly and the structure coding output by the encoder lacks corresponding constraints,it may lead to inaccurate structure coding,but in this study,it is only for personal images.Therefore,the generation of pose regression constraint structure code is added;Considering the failure of the constraint of the patch co-occurrence discriminator caused by the small proportion of the person region in the person image,a human object detection module is employed to segment the foreground and background of the generated person image and texture image,and texture constraints are made for the foreground and background respectively.The proposed models are evaluated on the Deep Fashion dataset,and the qualitative and quantitative results demonstrate their advantages,and the generated person images achieve advanced visual quality.