Design Of Virtual Try-on System Based On Generative Adversarial Network

The purpose of virtual try-on is to transfer the image of the target clothing to the image of the reference person,which has been a hot topic in recent years.The prior art usually focuses on preserving the original features of the clothing image on the generated image.However,when a large number of occlusions and complex poses appear in the reference person image,it is still a challenge to generate a clear and reasonable try-on image.In this thesis,a progressive generation logic is used to first generate a predicted semantic segmentation map,and then the predicted semantic segmentation map combines the original reference person and clothing to adaptively retain the non-target area,and generate the target area information to complete the construction of the try-on image.In addition,this thesis uses small datasets and small-size images for training,and also trains a super-resolution model based on a Generative Adversarial Networks to optimize the resolution of the generated try-on images.The main work of this thesis is as follows:（1）An Image-based Virtual Try-on Network by Adaptively Generating（VTNAG）is proposed.A unique repair module is used to fix the broken arm in the generated semantic segmentation map.The thesis adjust the structure of the generator in the network to make the image fusion more flexible.In the clothing deformation module,thin plate spline deformation is used,supplemented by mesh elastic constraints,which can effectively improve the situation of exaggerated deformation,and the clothing texture details of the generated images are clear.The effect of virtual try-on surpasses existing state-of-the-art models both qualitatively and quantitatively.（2）The Second-Order Degenerate Super-Resolution Network with U-Net Discriminator（SODSR）is proposed.In this thesis,the second-order degradation model will be used to synthesize image pairs for training,and ringing artifacts and overshoot artifacts will be added when synthesizing image pairs,so that the network can autonomously learn the law of image degradation in the real world.This thesis adopts the RRDB structure generator and the unique U-Net discriminator to learn the mapping rules between image pairs.In addition,this paper adds a spectral normalization layer to the U-Net discriminator,which effectively alleviates the convergence problem.The effect of super-resolution is qualitatively and quantitatively superior to existing state-of-the-art models.