Research On Virtual Try-On Based On Two-Dimensional Image

Virtual try-on allows consumers experience wearing different clothes without actually taking off and changing clothes.It is particularly important to bridge the gap between online clothing shopping and brick-and-mortar shopping.This this article conducts the thorough research of clothing alignment,try-on synthesis and network framework.Through deep learning technology,effective clothing alignment strategy and try-on synthesis model are proposed,and a virtual try-on scheme without clothing template is designed.The main work includes:（1）A clothing alignment module with double warping combining pixel warping and feature warping is proposed.Considering that the geometric matching method with thin-plate spline transformation can preserve the details but cannot accurately align with the reference person,while the method based on appearance flow can achieve better alignment effect but cannot preserve the full details.Therefore,this paper proposes a double warping strategy,which first warps the pixels of the target clothes and then warps the features,so that the warped clothes not only retain the full details of the clothes,but also have the effect of accurate alignment.（2）A synthesis model based on convolutional neural network block and Swin-transformer block is proposed,which can extract and integrate global and local features.Considering that the convolutional neural network has the characteristics of inductive bias but it also has the inherent limitation of limited receptive field,the Transformer has the ability of global modeling but it also needs large-scale training data.This paper proposes a synthesis model based on convolutional neural network and transformer to improve the overall network performance and generate high authenticity try-on results.（3）A toward 2D image-based user-friendly virtual try-on scheme is proposed.Considering that it is too difficult to obtain the target clothing template,this paper constructs a triple dataset by exchanging the clothes on two persons,designs a three-step training process including coarse synthesis,clothing alignment and refinement synthesis,and realizes a virtual try-on network framework without using the target clothing template.Experiments on the public dataset show that the scheme proposed in this paper has been greatly improved in both qualitative comparison and quantitative analysis.Compared with the existing methods,this method generates highly-realistic try-on images while retaining the clothing details and identity information.