面向开放场景虚拟试衣的服装图像规范化复原  

Normalized restoration of clothing images for virtual try-on oriented towards open scenes

作　　者：王志城 黄荣 董爱华[1,2] 王直杰[1,2] WANG Zhicheng;HUANG Rong;DONG Aihua;WANG Zhijie(College of Information Science and Technology,Ministry of Education,Donghua University,Shanghai,China;Engineering Research Center of Digitized Textile&Apparel Technology,Ministry of Education,Donghua University,Shanghai,China)

机构地区：[1]东华大学信息科学与技术学院,上海 [2]东华大学数字化纺织服装技术教育部工程研究中心,上海

出　　处：《东华大学学报（自然科学版）》2024年第6期133-139,共7页Journal of Donghua University(Natural Science)

基　　金：国家自然科学基金(62001099);中央高校基本科研业务费专项资金(2232023D-30)。

摘　　要：基于深度学习的虚拟试衣模型难以直接应用于从开放场景中获取的带有扭曲的非规范化服装图像,提出一种面向开放场景虚拟试衣的服装图像规范化复原网络,对带有扭曲的服装图像进行规范化复原,弥补非规范化服装图像与现有虚拟试衣模型之间的适配鸿沟。所提出的服装图像规范化复原网络以U-Net为骨干,以图到图的形式生成规范化复原图像。在U-Net的跳跃连接中设计了服装特征图扭曲模块和软门控单元,分别实现了特征对齐和区域筛选,有利于提高规范化复原图像的视觉质量。在规范化复原和虚拟试衣两个方面验证了所提模型的有效性。试验结果表明,所提模型的定量指标SSIM(structural similarity index measure)、PSNR(peak signal-to-noise ratio)、LPIPS(learned perceptual image patch similarity)分别达0.771、22.413、0.213,并在规范化复原图像中较好地保留了服装的图案及纹理,能够满足开放场景的虚拟试衣需求。The existing virtual try-on models based on deep learning are difficult to be applied to the distorted non-normalized clothing images directly obtained from open scenes.To solve this problem,a clothing image normalized restoration network is proposed for virtual try-on oriented towards open scenes,which normalizes the distorted clothing images and bridges the adaptation gap between the non-normalized clothing image and the existing virtual try-on models.The proposed clothing image normalized restoration network uses U-Net as the backbone to generate normalized restoration results in an image-toimage manner.A clothing feature map warping module and a soft-gated unit are designed in the skip connection of U-Net to realize feature alignment and region filtering,respectively,both of which are conducive to improving the visual quality of normalized restoration results.Experimental results of normalized restoration and virtual try-on verify the effectiveness of the clothing image normalized restoration network.The experimental results show that the proposed model achieves 0.771,22.413 and 0.213 in structural similarity index measure(SSIM),peak signal-to-noise ratio(PSNR)and learned perceptual image patch similarity(LPIPS),respectively,and the patterns and textures are well preserved in the normalized restoration results.These achievements can meet the needs of virtual try-on in the open scenes.

关 键 词：虚拟试衣 图像复原 U-Net 特征图扭曲 软门控 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]