【作者】 刘江；

【导师】 张淑芳；

【作者基本信息】 天津大学 ， 信息与通信工程， 2021， 硕士

【摘要】 近年来,随着互联网的飞速发展,催生了线上购物这一新型消费模式。线上购物中服装类产品占据主导地位,然而线上购物体验较差,缺乏一个接近现实的购物环境。消费者无法试穿线上服装,这导致网购的退货率和投诉率远高于线下实体店。为了解决线上服饰无法试穿这一问题,虚拟试衣技术应运而生。基于上述背景,本文就虚拟试衣应用展开了研究。主要工作可以分为四大部分:第一,针对原始基于无参数化模板的三维人体重建方法的局限性,本文提出了基于关键点的密集采样策略和多视角协同训练策略,通过增加额外的监督信息使得模型学习到更多细节特征,相比于经典算法,本文方法在脸部、手部和侧面体型等细节特征重构上有了明显提升;第二,针对基于参数化人体模型的三维人体重建方法,本文提出了基于姿态先验估计的多视角迭代配准策略,该策略通过一个预处理模型获得姿态先验值,然后通过多视角迭代拟合优化对应的参数值,从而加速模型收敛,并且模型能够充分考虑不同视角下的轮廓姿态信息,所重建的人体模型更加接近真实人体;第三,将参数化人体模型注册到密集隐式函数网络的输出模型上,注册后的模型既保留丰富的个性化细节特征,同时也具有灵活可控性。最后结合数字服装库实现3D虚拟试衣应用,并且从不同角度验证了本文的虚拟试衣算法的鲁棒性,实验效果较好,对于未来的虚拟试衣存在一定的应用价值。更多还原

【Abstract】 In recent years,with the rapid development of the Internet,online shopping,a new consumption mode,has been born.Clothing products dominate online shopping,but online shopping experience is poor and lacks a realistic shopping environment.Consumers cannot try on online clothes,which leads to the return rate and complaint rate of online shopping is much higher than that of offline physical stores.In order to solve the problem that online clothes can not be tried on,virtual fitting technology came into being.Based on the above background,we have conducted research on the application of virtual fitting.The main work can be divided into four parts: First,in view of the limitations of the original 3D human body reconstruction method based on nonparameterized templates,we propose a dense sampling strategy based on key points and a multi-view collaborative training strategy.By adding additional supervision information,the model can learn more detailed features.Compared with the classical algorithm,the proposed algorithm has a significant improvement in the reconstruction of detailed features such as face,hand,and side body shape.Second,for the threedimensional human body reconstruction method based on the parameterized human body model,we propose a multi-view iterative registration strategy based on posture prior estimation.This strategy obtains the attitude prior value through a preprocessing model,and then optimizes the corresponding parameter value through multi-view iterative fitting,thereby accelerating the model convergence.Since the model can fully consider the contour posture information from different perspectives,the reconstructed human body model is closer to the real human body;Third,the parametric human body model is registered to the output model of the dense implicit function network.The registered model retains rich personalized details and is also flexible and controllable.Finally,the 3D virtual fitting application is realized by combining the digital clothing library,and the robustness of the virtual fitting algorithm in this thesis is verified from different angles.The experimental effect is good,and there is a certain application value for the virtual fitting in the future.更多还原