Reasearch And Application Of Generative Adversarial Network Based On Feature Preaservation In Arbitrary Virtual Try-on System

In modern life,with the vigorous development of clothing e-commerce and the exposure of big data,more and more offline clothing merchants have begun to use e-commerce to sell clothing by displaying clothing pictures on the e-commerce platforms,however,users often encounter this question when shopping online: what will this dress look like on themselves? Therefore,to solve this problem,virtual try-on technology can help users choose their favorite clothing products.As a result,how to use massive clothing image data and advanced algorithms to develop virtual try-on method to enhance users' desire of purchasing and shopping experience is an important issue among clothing e-commerce platforms.Motivated by the recent research of virtual try-on,this paper applies the generative adversarial network with feature preservation to arbitrary virtual try-on system.The main work is as follows:1.A new dataset is proposed,which contains pairs of clothing pictures and model pictures of tops,bottoms,and whole outfits.Compared with the existing benchmark dataset,the new dataset includes additional pictures of the bottom clothes and whole outfits both male and female models,making it more diversified.In addition,more shapes and movements of the models are contained in the new dataset making it more complex.2.To handle cross-categorical virtual try-on,the limb prediction model based on deep convolutional neural network is proposed.The network uses the U-Net network structure and feature correlation calculation to coarse-grainy predict the limb skin exposed by the human body after changing clothes,and transfer the prediction results to the subsequent model group.Among them,U-Net network can retain human body related information,and the feature correlation calculation can combine clothing and human body information.This model solves the problem of limb prediction in different categories of clothing try-on(such as long sleeves (?) short sleeves,trousers(?) shorts,etc.).3.A clothing elastic transformation model based on Wendland is proposed to realize the local elastic transformation of the picture,which can reduce the interference between different control points,and achieve the effect of reasonable distortion of the clothing in the picture.The entire model predicts the relevant transformation parameters of the radial basis function through the network,and then obtains the corresponding coordinate grid according to the parameters,and finally obtains the warped clothes through bilinear interpolation.Compared with the traditional TPS,this method can retain the characteristics of the clothing and improve the overall virtual try-on performance.4.A feature preservation model is proposed to adjust the features of the human body and clothing,and output the final dressing result.This model uses parallel U-Net structure,and U-Net shares high-level semantic information through Residual Block to achieve feature preservation.The two U-Nets respectively predict the mask and the rendered human body,and finally obtain the fusion result between the warped clothing and the rendered human body through mask weighting.This model effectively improves the problem of artifacts,making the result more realistic.