Research On Virtual Try-on Method Oriented Towards Open Scenes

With the popularity of clothing e-commerce,consumers’ demand for online shopping experiences is constantly increasing.Virtual try-on allows consumers to remotely and conveniently view the try-on effect when shopping online,improving their shopping experience while reducing merchant costs.Consumers have a wide range of application needs to take photos directly from open scenes to obtain clothing of interest and view try-on effects.However,existing virtual try-on methods based on image generation have the following two problems.Firstly,it is difficult to adapt to non-normalized clothing images with distortion obtained from open scenes,resulting in an adaptation gap problem.Secondly,the differences between reconstruction and tryon tasks were overlooked.To alleviate these two issues,this thesis focuses on researching virtual try-on methods for open scenes.A generative clothing image normalized restoration method based on U-Net is proposed to address the adaptation gap issue,and a mapping from non-normalized to normalized of clothing images is established.A clothing feature map warping module and a soft-gated unit were designed in the U-Net skip connection to achieve feature alignment and region filtering,respectively,which is conducive to improving the visual quality of normalized restored clothing results.The effectiveness of the proposed method was verified in two aspects: clothing images normalized restoration and virtual try-on.The experiments show that the proposed method can generate realistic normalized restoration clothing result images,and well preserve the patterns and textures of the clothing,bridging the adaptation gap between non-normalized clothing images and virtual try-on methods.A virtual try-on method based on cycle-consistency and dual-task driven is proposed to address the differences between reconstruction and try-on tasks.This method uses human parsing map as a bridge to measure the cycle-consistency of human layout before and after try-on,providing a supervisory signal driving network training for reconstruction and try-on tasks.On this basis,a method of jointly driving training with reconstruction and try-on dual-task was designed to guide the virtual try-on model network to adapt to the differences between the dualtask,thereby improving the try-on effect of cross attribute(such as sleeve length and collar height)clothing.The experiments of performance comparison,ablation,and visualization show that the proposed method effectively improves the visual effect of try-on,and outperforms existing virtual try-on methods in quantitative and qualitative results.The proposed clothing image normalized restoration method is cascaded with the try-on method can meet the requirements of virtual try-on oriented towards open scenes.Based on the above achievements,a "Try-on Magic Mirror" application system has been designed and built.This system integrates functions such as human parsing,clothing recommendation,and virtual try-on,achieving pixel level human body parsing,intelligent clothing recommendation,and digital virtual try-on for free users in open scenarios.To achieve the above functions,use C# to design a user interaction interface,Python and Python to build the backend functions.After testing,the "Try-on Magic Mirror" application system has practicality.