Researches On Virtual Try-on System Based On Semantical Information And GANs

With the rapid development of Internet technology,consumers can buy clothes online without leaving home.The proportion of clothing sales in e-commerce is large.Under the circumstance of the e-commerce,consumer will buy various clothing which they like,at the same time,businesses will provide the clothing model photos for reference,but customer can only see the pictures of models wearing the dress,they can't try it on like they do offline.This disadvantage will reduce the consumers' shopping experience.Virtual try-on technology presents users' virtual try on scene through algorithm model,which is helpful for online sales of clothing industry.Because of the disadvantages of high cost,complex model and hardware dependence,the traditional 3D fitting technology is not suitable for e-commerce scenario.Therefore,this paper mainly conducts the following research based on the virtual try-on technology under 2d image:(1)Human body image representation and semantic synthesis algorithm,we introduce the virtual human body image representation mode of the try-on mission to get the body posture and semantic mapping segmentation algorithm,and proposes semantic and segmentation map layout synthesis algorithm based on the target clothes,the experiment confirms the validity of the proposed method.(2)Clothing warping algorithm,clothing deformation algorithm based on spatial transform network can make any style of clothes and body image region match,and put forward the garment warping network and uses the deformable convolution kernels and parameter regularization to improve the image quality.Qualitative analysis and quantitative result comparison shows the validity of the proposed algorithm.(3)GAN based feature fusion network,this chapter proposes a feature fusion network based on conditional GAN.Aiming at resolving semantic crossover,feature entanglement,image blur and other problems in the testing process of existing methods,a semantic spatially denormalization optimization strategy is proposed,and the ablation experiment is carried out to demonstrate its effectiveness.Qualitative and quantitative experiments show that the proposed method is superior to the existing feature fusion algorithms.(4)A virtual try on software is designed and developed,and a lightweight Streamlit library is used to deploy the virtual try on algorithm model.We introduce the process of the virtual try on task and shows the virtual try on sample.