Application Of Deep Learning Algorithm In Virtual Try-on

With the development of deep learning in recent years,image-based virtual try-on technology has gained more and more attention.At present,there are two main technologies for implementing virtual try-on using deep learning algorithms.One is based on CAGAN,but the image quality generated by the network is not ideal and cannot handle large spatial deformations.The other is based on VITON,however,there are serious color distortion in the image before and after try-on,the generation effect is not ideal in the face of self-occlusion,and there are some defects in the deformation network.Therefore,this paper studies and solves these problems in the application of deep learning algorithms in virtual try-on.The main contributions of this paper are as follows:(1)A new clothing warping module is proposed.The input of this module are clothes and human body posture map,and the output is the clothes after warping.In order to further improve the warping effect of the network,this paper discarded the original method of extracting the entire human body pose as the input,and only extracted the pose of the area that needs to be replaced,which can reduce the influence of irrelevant poses on the warping.At the same time,mask loss is added to make the shape of the distorted clothes more accurate.Lastly,in order to ensure that the warped clothes can retain as much of the texture characteristics of the original clothes as possible,distance constraint is applied to the control points of the regression network of the module based on the two-stage training.A large number of experimental results on VITON,MPV and other datasets verify that the new module can produce better deformation effects.(2)A virtual try-on method based on improved CAGAN is proposed.The main body of the original CAGAN network is a generator similar to U-Net,but the network structure is so simple.Therefore,in this paper,a multi-scale generator network is proposed based on CAGAN.The whole model consists of multiple networks,Firstly,the first network is used to generate coarse try-on results.Secondly,the second network is used to deal with the inaccuracy of the initial mask generated by the first network and deal with self-occlusion.Finally,the clothes after the new warping module are combined with the results of the first two steps to obtain the final try-on results.A large number of experimental results on datasets such as VITON have verified that this method can effectively solve the problems of the previous method and achieve better try-on results.(3)A virtual try-on method based on posture guidance is proposed.The previous methods mainly focus on fixed poses,and cannot complete the task of synthesizing images of people in arbitrary poses and clothing.Therefore,in order to further improve the try-on effect,this paper propose a multi-stage virtual try-on network structure based on VITON and introduce adversarial loss in the network.Firstly,the original character image and the target pose are taken as input,and the pose segmentation network is used to obtain the segmentation map of the character with the target pose.Secondly,the pose transformation of the character is performed by the pose transformation network to obtain the coarse try-on result.Finally,the clothes after the new warping module and the coarse try-on result are further refined and rendered.The mask between the warped clothes and the coarse results is learned,restore texture details and remove some artifacts and get the final try-on result.A large number of experimental results on MPV and other datasets verify that this method can achieve very good try-on results.