结构外观度量下的无监督人体图像生成  

Unsupervised Human Image Generation Based on the Appearance Consistency Loss

作　　者：赵应丁[1] 何俞玲 杨文姬[1] 吴沧海[1] 杨红云[1,3] 黄丽芳 ZHAO Ying-ding;HE Yu-ling;YANG Wen-ji;WU Cang-hai;YANG Hong-yun;HUANG Li-fang(Schoolof Software,Jiangxi Agricultural University,Nanchang 330045,China;School of Computer and Information Engineering,Jiangxi Agricultural University,Nanchang 330045,China;Key Laboratory of Agricultural Information Technology of Colleges and Universities in Jiangxi Province,Nanchang 330045,China;Jiangling Holdings Limited,Nanchang 330052,China)

机构地区：[1]江西农业大学软件学院,南昌330045 [2]江西农业大学计算机与信息工程学院,南昌330045 [3]江西省高等学校农业信息技术重点实验室,南昌330045 [4]江铃控股有限公司,南昌330052

出　　处：《小型微型计算机系统》2021年第10期2101-2106,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61462038,61562039,61966016)资助;江西省教育厅科技项目(GJJ190217,GJJ190180,GJJ200428)资助;浙江大学CAD&CG国家重点实验室开放课题项目(A2029)资助.

摘　　要：针对目标姿态下的人体图像生成过程中需要成对数据的局限性问题,提出了基于CycleGAN循环一致性思想的无监督生成方法.首先由源图像生成目标姿态图像,再由生成的图像映射回源图像.通过两次对抗训练优化生成器参数,使得生成的目标图像趋于逼真的效果,避免了成对数据的需求,可以应用到其他非刚性对象的转换任务中.此外,为了进一步地加强生成图像的特征一致性和结构一致性,本文提出了新颖的外观一致性损失.最后,在DeepFashion和Market-1501数据集上对提出的方法进行了实验,并与其它先进的模型进行了比较.结果表明从定性和定量两方面都取得了满意的结果.In order to solve the limitation of the need for paired data in generating person image under target pose,we propose an unsupervised generative adversarial network based on the cycle consistency idea of CycleGAN.We use a bidirectional generator to map back the generated images to source images,avoiding the need for paired data,and can be applied to other non-rigid object conversion tasks.In addition,in order to further enhance the feature consistency and structural consistency of the generated images,this paper proposes a novel appearance consistency loss.Finally,experiments were conducted on the DeepFashion and Market-1501 datasets and compared with other advanced models.The results showthat satisfactory results have been obtained from both qualitative and quantitative aspects.

关 键 词：图像生成 姿态转换 无监督学习 生成对抗网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]