Design And Implementation Of Posture Transfer Algorithm Based On Feature Merge And Texture Transfer Semantics

With the vigorous development of computer technology,artificial intelligence technology dominated by deep learning is getting closer to people’s lives.Artificial intelligence technology not only plays a huge role in the traditional security field and the field of face recognition,but also has also gradually entered the field of multimedia social networking.Among them,posture transfer is one of the popular directions,which has attracted the attention of a large number of researchers.The main goal of the posture transfer task is to give an initial posture image and key point information of the target posture,and uses the generative adversarial network to generate a human image corresponding to the target posture.The generated target posture image is required to be more samiar in the human posture with the target posture and in the texture with the initial posture image.In recent years,with the continuous advancement of deep learning technology,the generation effect of pose transfer has also been continuously improved,but the final generation effect is still not satisfactory,especially in the target pose and texture generation.There is a lot of room for improvement.In view of the challenges and problems in attitude transfer,the main research points of this article are as follows:(1)In view of the existing posture transfer algorithms,there are problems that the posture actions of the generated images are inconsistent with the target posture information,and the generated image texture information is fuzzy.This paper proposes a new type pose transfer algorithm based on multi-scale semantic transfer network.The network contains two modules,sementic parsing generation module(SPGM)and multi-scale semantic transfer module(MSTM).The sementic parsing generation module based on Encode-Decode network,generates the target sementic parsing mask guided by the target pose information to improve the similarity of generated posture.The multi-scale semantic transfer module extracts the features of the initial image,the generated semantic mask,the initial pose information and the target pose information in the different scales.Then the module transfers the pose in the different scales and merges the features by down-sampling pyramid to improve the similarity of texture.In addition,in order to improve the texture diversity of the generated image,the perceptual loss is added to the loss function to improve the generation of high-frequency features.Experiment on the dataset proves that the pose transfer algorithm based on the multi-scale semantic transfer network can generate images that are more similar to the target image in pose actions and textures.(2)After the image is down-sampled by convolution and then up-sampled by deconvolution,it will lose complex texture features such as patterns on clothes.To solve the problem of the original image texture pattern lost in the generated image,this paper proposes a new type of human image generation algorithm based on neighbor texture migration to improve the similarity between the texture of the generated image and the texture of the initial posture image.The algorithm uses the dense pose estimation algorithm to map the texture information of the initial image to the UVMap according to the body part,and after the target pose image is generated,the texture information of the extracted body part is remapped to the generated pose image.Experiments on the dataset prove that the image generated by the pose transfer algorithm based on the neighbor texture transfer has better retention of the complex texture information in the initial pose image.(3)In view of the current inconvenience of model deployment,this paper designs and implements a posture transfer system.The system is composed of inference interface module,inference service module and service scheduling module.The inference interface module is responsible for user login and upload and save data functions.The inference service module is responsible for receiving data for reasoning and other functions.The service scheduling module is responsible for accepting task messages from the inference interface module,processing the reasoning service applications without reasoning tasks offline and releasing the occupied computing resources,scheduling the reasoning service applications that are not online according to the message content,and forwarding the messages to the corresponding reasoning service application for processing.Proved by function test and performance test,the system can easily deploy well-trained pose transfer modules and deploy multiple models with limited resources.