Research On Human Movement Transfer Technology Based On Generative Adversarial Network

Human movement transfer has great application value and potential in the fields of character movement reproduction,virtual fitting,character animation,film and game production.The current action transfer model is mainly implemented by deep learning methods,which require the application of human pose estimation model for pose detection,and the application of generative adversarial network for image synthesis.Through the analysis and summary of the existing research,we found that these motion transfer models have two main problems: First,in the posture detection stage of the source and target characters,The existing human pose estimation models generally have some disadvantages,such as large amount of parameters and calculation,high redundancy,and long detection time,which affect the efficiency of the whole process of motion transfer;secondly,the quality of the action transfer image based on the traditional generation adversarial network is poor,resulting in the generated image picture is not real enough,the characters are deformed,The synthesized action of the target character cannot match the action of the source character well,which affects the final transfer effect.For the purpose of achieving accurate and fast human key point detection in the pose detection stage,this paper combines the Open Pose detection model and the Mobile Net V3 lightweight network to propose a human pose estimation algorithm MV3-CBAM-Open Pose based on a lightweight attention mechanism.First,the lightweight network Mobile Net V3 is used to replace the original Open Pose backbone network VGG-19;second,the structure of Open Pose's two-branch multi-stage convolutional neural network is modified to incorporate other network layers except the output stage;finally,the attention mechanism module CBAM,which combines space and channel,is introduced to balance the speed and accuracy of the model.Experimental results show that the combination of a lightweight network and an attention mechanism can effectively reduce the scale of the model while maintaining a high detection accuracy and recall rate,and improve the detection speed of the model.In order to generate higher quality motion migration images,we have fully studied the generative adversarial network derivative model pix2 pix HD,and proposed an improved pix2 pix HD.The network structure of the generator and discriminator of pix2 pix HD was improved,and the loss function was optimized.The encoder and decoder architectures are used in the generator network,and through the jump connection between the codecs,the deep and shallow feature information in the network can be effectively merged;4 4 convolution kernel are used in the discriminator network to replace 3 3 convolution kernel,the reason is that even-numbered convolution kernel combined with a step size set to 2 can avoid the checkerboard effect;the standard deviation matching loss is introduced on the basis of the loss function of pix2 pix HD to stabilize the training of the GAN model.Through experimental verification,the improved pix2 pix HD produces better image quality.Combining the proposed human pose estimation algorithm MV3-CBAM-Open Pose and the improved pix2 pix HD,a new motion transfer model is proposed.This article comprehensively compares three types of motion migration models: EDN(Everybody Dance Now),NKN(Neural Kinematic Networks),and LCM(Learning Character-Agnostic Motion),and the proposed action migration model was superior to the proposed one in terms of visual effects and quantitative indicators.