Research On Human Image Generation Model Based On Structure Appearance CycleGAN And Graph Convolutional Network

In recent years,with the development of online fitting,intelligent transportation and other technologies,the demand for human image data is also increasing sharply.Human image generation model has become a critical research topic.However,in the face of complex background,uneven image distribution and high requirements for detail texture,how to generate a real human image with clear image and consistent with human perception in all aspects is still a challenging problem.Compared with the traditional generation model,the generation countermeasure network has been proved in recent years.In some mainstream tasks such as image generation,it has more advantages in generation ability,model generalization ability and robustness.For improving the quality of human image generation,new human image generation models are proposed.Aiming at the limitation of pairwise data in the process of human image generation under target posture,an unsupervised generation model based on cyclegan cyclic consistency is proposed in this paper.Through a set of mirrored generators sharing parameters,the generation effect of the generator is constrained by remapping the generated image back to the source domain without using paired data sets.The model avoids the need of paired data,can be applied to the transformation tasks of other non rigid objects,and improves the generalization ability of the model.In order to further improve the image quality,during the model training process,the appearance consistency loss is used to extract the high-level features of the source image and the generated source image by using the VGG19 network,and further improve the generator in terms of texture details and pose structure.Experiments show that the model improves the image quality of unsupervised human image generation.Aiming at the problem of non rigid deformation and complex entanglement in the conversion from source attitude to target attitude,the traditional convolution neural network needs to overlay multiple convolution layers to obtain the global relationship between all joint points of source attitude and target attitude.This method will reduce the operation efficiency of the model.For settling the above above-mentioned matters,this paper proposes an attitude conversion method based on graph convolution,constructs an interaction space,maps the joint points of source attitude and target attitude into the interaction space respectively,creates a fully connected graph linking all joint points in the interaction space,and infers the global relationship of the graph.After reasoning,the updated joint points are remapped back to the original space to obtain the updated pose code.Experiments show that this method effectively improves the structural consistency of the generated image.In view of the unclear texture details generated by human body image,this paper proposes to decompose the source image into different attributes(such as hair,coat,trousers and shoes)by using the human body analysis model based on VGG network,and then input the decomposed attributes into the appearance texture encoder to reconstruct the appearance coding.Through the combination of the reconstructed appearance coding and the updated posture coding,Generate more realistic human body images.This paper adopts the strategy of alternating update to promote the mutual guidance between the pose module and the appearance module,and gradually generate the target human body image.To prove the validity of the models,we conducted experiments on Deep Fashion data set and Market1501 data set,and compared with other advanced models in both qualitative and quantitative aspects.The results show that the details of the character image generated by the proposed model are more delicate,more real,and less different from the target pose.