基于人体语义分割重构的虚拟试衣网络

Virtual Try-on Network by Reconstructing Human Semantic Segmentation

近年来,虚拟试衣在网上购物中发挥重要作用,其商业价值被广泛关注。基于图像的虚拟试衣旨在将目标衣物准确形变,并合成到目标人物图像上。而实现衣服准确形变的关键在于如何确定衣服图像的形变位置和形变程度。针对该问题,提出一种基于人体语义分割重构的虚拟试衣网络。在虚拟试衣前,首先利用仿射变换将目标衣物预对齐到目标人体的相应位置;然后重构目标人体的语义分割数据,预测虚拟试衣成功后的语义分布;最后根据新的语义分割计算形变参数,并形变衣物图像以自适应补充缺失姿态信息的方式合成到目标人体上。实验表明,该方法可以取得更好、更真实的虚拟试衣结果。

In recent years,virtual try-on technology has begun to play an important role in online shopping,and its commercial value is also widely concerned.Image-based virtual try-on is designed to accurately transform the target clothing and synthesize it onto the target image.The key to realize the accurate deformation of clothes is how to determine the deformation position and deformation degree of clothes image.To solve this problem,this paper proposes a virtual try-on network based on human semantic segmentation and reconstruction.Before the virtual try-on,the target clothing is firstly prealigned to the corresponding position of the target human body by affine transformation.Then,the semantic segmentation data of the target body is reconstructed to predict the semantic distribution after the successful virtual try-on.Finally,the deformation parameters are calculated according to the new semantic segmentation,and the deformed clothing image is synthesized to the target body in a way that adaptively fills the missing pose information.Experiments show that this method can achieve better and more realistic results of virtual try-on.